Engram: Cognitive Memory Layer for AI Agents

The SQLite of agent memory: embeddable, local-first, cognitively grounded.

Engram architecture: observations flow into the memory core, which holds episodic memory, bitemporal semantic facts, and an entity graph, with optional LLM reflection and compression, and recall flowing back out to the agent

Engram gives an AI agent a persistent, provenance-tracked memory: raw episodic observations, bitemporal semantic facts that know when they were true, and an entity graph for associative recall, all in one embeddable .engram file with no server, no Docker, and no API key required to write a memory.

The problem

Every AI agent starts from zero. Ask it something it answered last week - it has no idea. Show it a document it already processed - it processes it again. Tell it Ivan moved to a new company - it still thinks Ivan works at the old one.

This happens because agents have no persistent memory. When the conversation ends, everything is gone.

The usual fix is to throw a vector database at the problem. Store text, embed it, search by similarity. That helps - but it's not enough. You still can't ask "what did the agent think in March?" or "where did this belief come from?" or "show me everything the agent knows about Ivan." A vector search finds similar text. It doesn't understand time, relationships, or importance.

Engram is memory done properly.

What Engram does

Engram gives your agent a persistent memory that works like a file - one .engram file on disk, no server required. You pip install it and start using it in two lines:

from engram import Engram

with Engram(path="./agent.engram") as mem:
    # Remember something
    mem.observe("Ivan moved from Acme to Globex last week", actors=["Ivan"])

    # Recall it later - even in a completely different session
    for r in mem.recall("where does Ivan work?", k=3):
        print(f"[{r.score:.2f}] {r.episode.content}")

No server to start. No API key for the store. No Docker. No configuration file.

Here is what Engram gives you that a plain vector database does not:

Remembers raw events - every observation is stored with who was involved, what tags apply, and how important it felt at the time. Search finds the right memories even when the query is phrased differently.

Understands facts - a background process (no LLM needed at write time) reads your observations and extracts structured knowledge: Ivan works at Globex, Alice is the CTO. These facts can be queried directly, updated when things change, and traced back to their source.

Knows what happened when - if Ivan changes jobs, the old fact is not deleted. It is closed with an end date. You can ask what the agent believed in March even if the truth has changed since.

Forgets wisely - memories that haven't been accessed in a while gradually become less important. Memories that matter (accessed often, emotionally significant) stay sharp. The agent doesn't accumulate noise forever.

Explains itself - for any fact, you can ask where it came from: which observation triggered it, which LLM run extracted it, with what confidence.

Works with multiple agents - several agents can share a single .engram file. Each has its own private observations; extracted facts and the relationship graph are shared between them.

Engram architecture: an agent observes and recalls against one SQLite file holding episodic, semantic and procedural memory with bitemporal facts, spreading activation and decay, exposed over MCP on stdio

_{The same service as its room on it-rat.com draws it, where the diagram sits next to a simulation you can scrub back and forth.}

Where this fits in the stack

Engram is the memory plane of the TAIPANBOX agent-governance stack: it gives agents persistent, provenance-tracked memory, and its own reflection can route through TokenFuse.

flowchart TB
  Agent["AI agent (any framework)"] -->|"LLM call (base-URL swap)"| TF["TokenFuse proxy: spend + enforcement"]
  TF -->|"POST /v1/decide (PEP)"| WX["Wardryx: policy PDP"]
  WX -.->|"allow / deny / hold"| TF
  TF -->|"cheapest model, budget OK"| LLM[("LLM provider")]
  TF -->|"CallRecords"| CL["TokenFuse Cloud: control plane, incidents, replay, evidence, kill-switch"]
  VCX["Vouchryx: delegation proved, and endable"] -->|"short-lived token: act + cnf"| TF
  TF -.->|"polls /v1/revocations"| VCX
  VCX ==>|"delegation_issued / denied / revoked"| BUS
  TF ==>|"agent-event NDJSON"| BUS{{"agent-event bus + Agent Passport"}}
  WX ==> BUS
  Agent -->|"web fetch"| SCX["Scopyx: governed web egress"]
  SCX -->|"POST /v1/decide"| WX
  SCX ==>|"web_fetch / web_blocked"| BUS
  ENG["Engram: memory"] -->|"reflect via base_url"| TF
  ENG ==> BUS
  BUS ==> IDX["Idryx: identity graph, detectors, Agent-BOM"]
  IDX ==>|"identity_finding"| BUS
  BUS ==> QX["Qryx: crypto / PQC, passport + hash-chain scan"]
  QX ==>|"crypto events"| BUS
  BUS ==> VX["Verdryx: quality / drift"]
  VX ==>|"quality events"| BUS
  TF -->|"outcome-tagged traces"| VX
  MX["Mockryx: pre-prod safety rehearsal"] -->|"hostile scenarios"| TF
  MX ==>|"sim events"| BUS
  BILL[("cloud, SaaS and model bills")] --> CC["CostCrew: the bill, worked by a crew of agents"]
  CC ==>|"spend_spike / budget_threshold / crew moves"| BUS
  BUS ==> TRX["Trailryx: the record plane, sealed and packed"]
  BUS ==> HX["reads the log, mails you (heraldyx)"]
  HX -->|"one mail, a view and never an action"| OPS["your mailbox"]
  HX ==>|"alert_sent"| HJ[("heraldyx's own hash-chained journal, not this bus")]
  YOU(["you, in a browser over your own tunnel"]) --> GX[["Genaryx: the console over all of it"]]
  GX -->|"signed commands: the kill, an approval, a policy"| CL
  GX -->|"signed commands"| WX
  GX ==>|"console_command"| BUS
  GX -.->|"reads it"| IDX
  GX -.->|"reads it"| QX
  GX -.->|"reads it"| VX
  GX -.->|"reads it"| MX
  GX -.->|"reads it"| ENG
  GX -.->|"reads it"| SCX
  GX -.->|"reads it"| CC
  TFP["terraform-provider-taipan"] -->|"budgets + passports as code"| CL
  ASG[["agent-stack-go: shared Go contract"]] -.->|imported by| IDX
  ASG -.->|imported by| WX
  ASG -.->|imported by| MX
  ASG -.->|imported by| TFP
  ASG -.->|imported by| HX
  ASG -.->|imported by| QX
  SPEC[["agent-passport: the spec"]] -.->|governs| BUS

Consumes: agent observations (observe()).
Produces: source: engram events and why() provenance, opt-in via the Agent Passport agent-event envelope, schema taipanbox.dev/agent-event/v0.2.
Talks to: TokenFuse (reflection's LLM adapter can point base_url at TokenFuse); governed by the agent-passport agent:// scope.

The full stack is TokenFuse (spend), Wardryx (policy), Vouchryx (delegation), Engram (memory), Idryx (access), Qryx (crypto), Verdryx (quality), Mockryx (pre-prod), scopyx (governed web egress), CostCrew (the bill), Trailryx (the record) and heraldyx (the mail out), on the shared Agent Passport + agent-event contract (agent-stack-go / agent-passport), configured via terraform-provider-taipan and driven from Genaryx, the console over all of it.

Run the whole open stack locally with one command via stack-up; the stack's home on the web is it-rat.com.

Live infrastructure validation

Before any public launch, Engram's Anthropic adapter was run against a real Claude model on real Linux infrastructure: three independent runs, zero bugs, zero contradictions, every extracted fact carrying full why() provenance back to the observations it came from.

Full write-up and all numbers: VALIDATION.md.

Running it on a Kubernetes cluster

The whole stack was deployed as a five-node k3s cluster on Hetzner, AWS and GCP between 25 and 27 July 2026 (six clusters, all destroyed afterwards). The manifests, the traps and the evidence are public in stack-k8s. Engram is deliberately not a pod. The console speaks to engram-mcp over stdio, and a sidecar container cannot be another container's stdin, so the binary has to live inside the console image rather than beside it. That is why images/console.Dockerfile is a four-language build. Getting this wrong produces a cluster where the memory tab is permanently empty while every pod reports healthy. The store itself takes an ordinary ReadWriteOnce volume; single-file memory is an asset here, because there is nothing to cluster.

To be clear about scope: those runs verified the deployment shape and the service coming up correctly on three clouds. They did not exercise memory under load: the memory plane was deliberately left unseeded on those clusters, so there are no recall or reflection numbers from them, and none are claimed. The numbers in this file remain the ones to read.

What is Engram, technically?

The three memory kinds: episodic observations with vector and BM25 recall, bitemporal semantic facts, and the entity relationship graph

Engram is a cognitive memory layer for AI agents - a single local file (agent.engram) built on SQLite. It models three kinds of memory that mirror how human memory works:

Episodic memory - raw observations stored as they happen, with actors, tags, salience, and emotional weight. No LLM required at write time; writes complete in ~4 ms.

Semantic memory - structured knowledge extracted from episodes via a background reflection loop: (subject, predicate, object) triples with full bitemporal validity. Every fact tracks when it was true in reality and when the system learned it - independently on two timelines. When Ivan switches jobs, the old fact is closed with valid_to, not deleted. You can query what the agent believed in March even if the truth has since changed.

Dynamic importance - each memory carries a living importance score based on the Ebbinghaus forgetting curve, reinforced by retrieval frequency and emotional weight. Memories below threshold decay and are pruned automatically during reflection. The agent forgets what doesn't matter; critical memories survive.

What you can actually do with it

Debug beliefs: when the agent says "Ivan works at Globex," call mem.why(fact_id) to see exactly which episode produced that belief, which reflection run extracted it, which model, and with what confidence.
Erase a person: forget_entity("Ivan") permanently removes all episodes, facts, and graph edges connected to Ivan - a proper GDPR right-to-be-forgotten.
Query the past: mem.recall("Ivan employer", as_of=datetime(2024, 3, 1)) returns what the agent knew at that exact point in time, not what it knows now.
Run multiple agents: a planner and a coder can share one file - each sees its own episodes, both benefit from shared extracted facts.
Plug into any MCP client: run engram-mcp --db ./agent.engram and Claude Desktop, Claude Code, or Cursor can remember, recall, why, and forget against the same store with zero integration code - see MCP Server in the API reference.

What the memory actually holds

A vector store answers one question: what text looks like this text. Engram answers the questions an agent asks after that one.

Capability	What it is here
Hybrid recall	BM25 over FTS5 and cosine over sqlite-vec, blended, so an exact term and a paraphrase both find the episode. The blend was swept rather than guessed
Bitemporal facts	Every fact carries when it was true and when it was recorded, so `as_of` answers what was believed in March, not only what is believed now
Spreading activation	Recall walks Hebbian-weighted edges out from what matched, so a query reaches what the match is connected to and not only what it resembles
Importance and decay	An Ebbinghaus curve with Hebbian reinforcement: what is used stays sharp, what is untouched fades, and reflection prunes the bottom. The store stays useful instead of growing
Working memory	A 7±2 scratchpad, Miller's number, for what the agent is holding right now as opposed to what it knows
Provenance	`why()` answers where a belief came from, episode by episode, so a wrong answer can be traced to what produced it
Contradiction	Reflection finds facts that disagree and records the disagreement rather than overwriting one with the other
Erasure	`forget()` at episode, entity or fact level, and `compress()` burns the raw text on a schedule while the extracted facts stay, so a person can be removed without the history losing its shape
Many agents, one file	`agent_id` is a partition key: each agent keeps its own episodes, all of them share extracted facts, and a scoped recall stays flat as the store grows

Everything above is one file. pip install, open it in two lines, no server, no Docker, no API key for the store itself, and no LLM call on the write path: writes land immediately and reflection runs later, in the background, when you ask for it.

It speaks MCP natively, so Claude Code, Claude Desktop or Cursor can remember, recall, why and forget against the same file with no integration code, and it plugs into LangChain and LlamaIndex where those are already in use.

How Engram works

Memory that doesn't forget the wrong things

Most tools either remember everything forever (noise accumulates) or forget everything when the session ends (nothing persists). Engram does neither.

Every memory gets an importance score. Memories you access often, or that carry emotional weight, stay sharp. Memories that sit untouched gradually fade. When the agent runs its background reflection pass, low-importance memories are pruned automatically. The result is a store that stays useful instead of bloating.

This is modelled on the Ebbinghaus forgetting curve - the same pattern that describes how humans forget - combined with Hebbian reinforcement from repeated retrieval.

Facts that know when they were true

When you just store text and search it, you lose track of time. "Ivan works at Acme" and "Ivan works at Globex" are just two strings - you don't know which is current, or what changed.

Engram extracts structured facts from your observations - triples like (Ivan, works_at, Globex) - and tracks two independent timelines for each:

When it was true in reality (valid_from / valid_to)
When the system learned it (recorded_at / superseded_at)

When Ivan changes jobs, the old fact is not deleted - it is closed with an end date. The new fact is added alongside it. You can query what the agent believed at any point in the past:

# What did the agent think about Ivan's employer in March?
mem.recall("Ivan employer", k=5, as_of=datetime(2024, 3, 1, tzinfo=UTC))

# Full fact history - every job Ivan ever had, with dates
mem.timeline("Ivan")

This two-timeline approach is standard in financial databases and audit systems. In the AI memory space, Engram is the only tool that implements it.

What a recall actually does

A recall runs two retrievers over one store: the query is embedded and searched by exact vec0 KNN with agent_id as a partition key and ts as a metadata column, and in parallel tokenised and searched by FTS5 BM25 with its terms joined by OR; each pool is min-max normalised and blended 0.5 cosine to 0.5 BM25 into the top k episodes

Two things in that picture are worth spelling out, because both were wrong until recently and neither is visible from the outside.

The filters live inside the scan. agent_id is a vec0 partition key and the timestamp is a metadata column, so "this agent's episodes" and "as of March" are conditions the index applies while it searches. They used to be applied afterwards, in a join, where a filter can only cut into a top-k that has already been chosen: in a shared store where one agent held most of the episodes, the other agent's recall(k=5) returned nothing at all rather than its own five nearest. It also means scoped recall no longer grows with the store. An agent holding five episodes pays for five, whether the shared file holds a thousand or a hundred thousand: 0.05 ms against 0.09 ms across that whole range.

The keyword side ORs its terms. FTS5 reads a space between words as an implicit AND, so joining a question's tokens with spaces demands that every one of them appear in a single episode. For "what was the name of the restaurant I mentioned when we talked about my anniversary dinner?" that is sixteen words; it matched nothing, and the blend quietly reduced to plain cosine. Hybrid mode was cosine under another name for any query longer than a few words. What exposed it was a benchmark run scoring both modes to three identical decimal places.

Three ways to search

Engram ships three retrieval modes behind the same API:

mode="hybrid" (default) - combines keyword search (BM25) with semantic search, then blends the scores. It is the default because on LongMemEval-S it beats pure vector search at every k measured, by most at the thing that matters to an agent: putting the specific turn that holds the answer in front of it (0.830 against 0.772 at k=5). Questions carry names, places and specifics, and lexical matching catches those exactly where an embedding blurs them into neighbours. It is also faster per query, 10 ms against 16 ms. The blend is configurable:

# BM25 keyword + cosine vector, weighted blend
results = mem.recall("Alice CTO Globex", k=5, mode="hybrid")

# More weight on exact keywords, less on semantics
results = mem.recall("quarterly budget", k=5, mode="hybrid",
                     vector_weight=0.3, fts_weight=0.7)

mode="cosine" - pure semantic vector search, no keyword component. Finds memories that mean the same thing as your query even when they share no words with it.

mode="spreading" - follows relationship edges between memories. If Ivan is connected to Project X in the graph, a query about Ivan can surface Project X episodes even if they share no words or meaning. One memory activates its associates, like human associative recall.

Technically: spreading activation runs BFS over Hebbian-weighted graph edges, ranking results by α·cosine_similarity + β·graph_activation + γ·importance_score.

How well it recalls, measured

All 500 questions of LongMemEval-S (ICLR 2025, MIT). Each question carries its own history of 30 to 60 sessions; every turn is ingested as one episode, 246 738 in total, and the question is asked against that store. No LLM anywhere in the loop.

mode	session@5	session@10	turn@5	turn@10	ms/query
`hybrid` (default, `0.5 / 0.5`)	0.970	0.982	0.830	0.894	10
`cosine`	0.956	0.978	0.772	0.862	16

The dataset marks both the sessions that hold the evidence and the individual turns, so there are two honest numbers, and the second one needs a caveat spelled out in the API reference: 59% of questions flag more than one evidence turn, so counting a hit when any of them is retrieved is an upper bound on what the model was handed. Session recall says the right conversation came back. Turn recall says one of the 896 flagged turns did, out of 246 738 - that is what the agent actually reads, and it is 14 points lower. A memory system quoting one unqualified "R@k" is quoting the first; ask which.

The blend behind hybrid was swept rather than assumed: every weighting from pure BM25 to pure vector was scored in the same pass, both ends came out worse than the middle, and the default is now 0.5 / 0.5, which leads the previous 0.7 / 0.3 on all four metrics by about five questions out of five hundred. The margin is small and said to be small; the shape of the curve is the part that matters.

Cosine reproduced its figures to three decimals across two independent six-hour passes over the whole dataset. Cosine is ahead on exactly one cut, multi-session at k=10 (0.992 against 0.985), which is in the API reference along with the per-type breakdown. The per-question records behind every number are in benchmarks/results/ with the snippet to recompute them, so the table can be checked rather than believed.

Reproduce it yourself:

engram-bench longmemeval --data ./longmemeval_s_cleaned.json --k 5,10 \
    --checkpoint ./lme.jsonl --resume

A scratchpad for the agent's current task

Engram also provides WorkingMemory - a small, fast, in-memory scratchpad for whatever the agent is actively thinking about. It holds a fixed number of items (default 7, matching the average human working memory capacity). When it fills up, the least-recently-used item is dropped - and if you pass an Engram store, it is automatically saved to long-term memory before being evicted:

from engram import WorkingMemory

wm = WorkingMemory(capacity=5, engram=mem)  # evicted items → long-term store
wm.set("task", "Summarise the quarterly report")
wm.set("context", "Revenue grew 12% YoY - needs explanation")

item = wm.get("task")      # read + promote to most-recently-used
item = wm.peek("context")  # read without changing eviction order

wm.flush()  # write everything to long-term store + clear

Background reflection (the agent's "sleep")

LLM calls in Engram never block writes. The reflection loop runs asynchronously - while the agent keeps working:

Group recent observations by entity or topic
Call the LLM to extract structured facts (Ivan works_at Globex)
Detect contradictions - same subject and predicate, different value
Close superseded facts with an end date (newest value wins; an identical re-extraction supersedes silently)
Recompute importance scores
Prune memories below threshold

thread = mem.reflect_async()  # starts in background, returns immediately
thread.join()                 # wait only when you need the results
print(f"{thread.result.facts_extracted} facts, {thread.result.cost_tokens} tokens")

Compressing old memories

When a store grows large, compress() groups low-importance observations into batches and asks the LLM to summarise each batch into a single paragraph. The originals are hard-deleted; the summary is stored in their place, with a summary_of pointer to what it replaced:

result = mem.compress(
    max_episodes=1000,        # only compress when store exceeds this
    importance_threshold=0.3, # target: episodes below this importance score
    batch_size=20,            # observations per LLM call
)
print(f"Removed {result.episodes_removed} episodes → {result.summaries_created} summaries")

Compression is lossy by design. Run reflect() first to extract facts from episodes before compressing them - facts survive compression, raw text does not.

Under the hood - technical details

Bitemporal validity

Every fact carries two independent timelines:

valid_from / valid_to       → when the fact was TRUE in reality
recorded_at / superseded_at → when the system LEARNED it

Hybrid BM25 + cosine recall

Three retrieval modes unified in one API:

mode="hybrid"    → FTS5 BM25 + cosine, normalised and blended (default)
mode="cosine"    → pure vector similarity (semantic)
mode="spreading" → cosine KNN seeds → BFS over Hebbian graph

Importance scoring formula

importance(m, t) =
    salience(m) × exp(−λ × (t − last_access(m)))   # Ebbinghaus forgetting curve
  + α × log(1 + access_count(m))                    # Hebbian reinforcement
  + β × emotional_weight(m)                          # affective weight

Parameters λ, α, β are configurable via DecayConfig.

Spreading-activation graph traversal

query → seed memories (cosine KNN)
              ↓
         graph edges (Hebbian weights - reinforced by co-access)
              ↓
         activated neighbors (activation × decay per hop)
              ↓
    rank by: α·similarity + β·activation + γ·importance

Working memory - Miller's 7±2 law

Fixed-capacity LRU cache backed by collections.OrderedDict. Evicted items optionally written to long-term store via observe(). Capacity default of 7 matches the average human working memory span (Miller, 1956).

Install

pip install engdbram

# LLM-powered reflection (optional, pick one):
pip install 'engdbram[anthropic]'   # Claude
pip install 'engdbram[openai]'      # OpenAI or any OpenAI-compatible

# Integrations:
pip install 'engdbram[mcp]'         # MCP server (Claude Desktop, Cursor, etc.)
pip install 'engdbram[langchain]'   # LangChain retriever + chat history
pip install 'engdbram[llamaindex]'  # LlamaIndex memory buffer

# Everything:
pip install 'engdbram[anthropic,mcp,langchain,llamaindex]'

The distribution name on PyPI is engdbram (the engram name is squatted). The import name is still engram, so application code is unaffected: from engram import Engram.

Requirements: Python 3.11+, no system dependencies. fastembed downloads the ONNX embedding model (~64 MB) on first use; all subsequent calls are local, including with the network unplugged.

Verify a download

Every release is signed keyless with Sigstore and carries a build-provenance attestation and an SBOM. With cosign and gh installed:

tag=<tag>
cosign verify-blob --bundle SHA256SUMS.sigstore.json \
  --certificate-identity "https://github.com/TAIPANBOX/engram/.github/workflows/release.yml@refs/tags/${tag}" \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com SHA256SUMS
sha256sum -c SHA256SUMS
gh attestation verify engdbram-<version>-py3-none-any.whl -R TAIPANBOX/engram

Releases through v2.4.1 have none of this; the tags say so.

Quickstart

Basic usage

from engram import Engram

mem = Engram(path="./agent.engram")  # or ":memory:" for ephemeral

# Store an observation - instant, no LLM needed
ep_id = mem.observe(
    "Alice presented the Q3 roadmap to the exec team",
    actors=["Alice"],
    tags=["work", "roadmap"],
    salience=0.8,           # 0-1, subjective importance at encoding
    emotional_valence=0.2,  # -1 (negative) … +1 (positive)
)

# Semantic recall
results = mem.recall("Alice roadmap", k=5)
for r in results:
    print(f"[score={r.score:.2f}] {r.episode.content}")

# Assert facts directly (no LLM)
mem.assert_fact("Ivan", "works_at", "Globex", confidence=0.95)

mem.close()

Async API

import asyncio
from engram import AsyncEngram, ObserveInput

async def main():
    async with AsyncEngram(path="./agent.engram") as mem:
        # All methods are async - event loop never blocked by ONNX or SQLite
        await mem.observe("Alice joined Globex as CTO", actors=["Alice"])
        await mem.observe_many([
            ObserveInput(content="Q3 planning complete", tags=["planning"]),
            ObserveInput(content="Ivan submitted architecture proposal", actors=["Ivan"]),
        ])

        results = await mem.recall("who joined Globex?", k=3)
        for r in results:
            print(f"[{r.score:.2f}] {r.episode.content}")

        await mem.assert_fact("Alice", "role", "CTO")
        facts = await mem.timeline("Alice")

asyncio.run(main())

Working memory scratchpad

from engram import Engram, WorkingMemory

with Engram(path="./agent.engram") as mem:
    # 5-slot scratchpad; evicted items automatically saved to long-term memory
    wm = WorkingMemory(capacity=5, engram=mem)

    wm.set("goal", "Draft the board presentation")
    wm.set("context", "Q3 revenue up 12%, but CAC increased")
    wm.set("constraint", "Must fit 10 slides, no more")

    task = wm.get("goal")        # promotes to most-recently-used
    note = wm.peek("constraint") # reads without changing LRU order

    print(f"Current slots: {len(wm)} / {wm.capacity}")
    wm.flush()  # write everything to long-term store + clear

Hybrid recall

with Engram(path="./agent.engram") as mem:
    # BM25 keyword match + cosine vector search, blended
    results = mem.recall("Alice quarterly roadmap", k=5, mode="hybrid")

    # Tune the blend weights
    results = mem.recall(
        "exact phrase match needed",
        k=5,
        mode="hybrid",
        vector_weight=0.3,  # less semantic
        fts_weight=0.7,     # more keyword
    )

Bulk import with observe_many

When loading historical context, observe_many() runs a single ONNX inference pass for the whole batch and commits all rows in one transaction - about 2× faster than calling observe() in a loop:

from engram import Engram, ObserveInput

items = [
    ObserveInput(
        content="Alice joined Globex as CTO",
        actors=["Alice"],
        tags=["hr"],
        salience=0.9,
    ),
    ObserveInput(content="Q3 planning session concluded", tags=["planning"]),
    ObserveInput(content="Ivan submitted the architecture proposal", actors=["Ivan"]),
]

with Engram(path="./agent.engram") as mem:
    ids = mem.observe_many(items)
    print(f"Inserted {len(ids)} episodes")

Async reflection with Claude

from engram import Engram, AnthropicAdapter

mem = Engram(
    path="./agent.engram",
    llm=AnthropicAdapter(model="claude-haiku-4-5-20251001"),
)

mem.observe("Ivan said he finally joined Globex last Monday")
mem.observe("The team shipped v2 of the payment service")

# Trigger reflection in the background
thread = mem.reflect_async()

# Keep doing agent work…
results = mem.recall("Ivan career", k=5)

thread.join()
run = thread.result
print(f"Facts: {run.facts_extracted}  Contradictions resolved: {run.contradictions_resolved}")
print(f"Tokens used: {run.cost_tokens}")

Memory compression

from engram import Engram, AnthropicAdapter

mem = Engram(
    path="./agent.engram",
    llm=AnthropicAdapter(model="claude-haiku-4-5-20251001"),
)

# Compress episodes with low importance into LLM summaries
result = mem.compress(
    max_episodes=500,         # no-op if store is smaller than this
    importance_threshold=0.3, # episodes below this score are candidates
    batch_size=20,            # episodes per LLM call
)
print(f"Compressed {result.episodes_removed} episodes → {result.summaries_created} summaries")
print(f"Tokens used: {result.cost_tokens}")

mem.close()

Time travel

from datetime import datetime, UTC

# What did the agent know about Ivan in March 2024?
past_results = mem.recall(
    "Ivan employer",
    k=5,
    as_of=datetime(2024, 3, 1, tzinfo=UTC),
)

# Full fact timeline for an entity
for fact in mem.timeline("Ivan"):
    end = fact.valid_to.date() if fact.valid_to else "now"
    print(f"[{fact.valid_from.date()} → {end}]  Ivan {fact.predicate} {fact.object}")

Multi-agent shared store

Multiple agents can read and write to the same .engram file. Episodes are scoped per agent; facts and the entity graph are shared.

from engram import Engram

# Each agent has its own episode scope. The id is an opaque key to the store,
# and an `agent://<trust-domain>/<name>` identifier is what the Agent Passport
# event log needs, so writing it that way from the start costs nothing.
planner = Engram(path="./team.engram", agent_id="agent://acme.example/planner")
coder   = Engram(path="./team.engram", agent_id="agent://acme.example/coder")

planner.observe("Decided to migrate to PostgreSQL", tags=["arch"])
coder.observe("Started migration branch: feat/pg-migration", tags=["dev"])

# Each agent recalls only its own episodes by default
planner_results = planner.recall("migration", k=5)

# Cross-agent search when needed
all_results = planner.recall("migration", k=10, cross_agent=True)

# Inspect who's written to the shared file
with Engram(path="./team.engram") as global_view:
    # ['agent://acme.example/coder', 'agent://acme.example/planner']
    print(global_view.list_agents())

planner.close()
coder.close()

Backup and export

# Hot backup - safe to call while the store is open. It copies the whole file,
# so it is refused on an instance scoped to an agent_id; open the file unscoped.
mem.backup("./agent_backup.engram")

# Portable JSON export (episodes, facts, entities, edges). A scoped instance
# exports its own episodes and edges; facts and entities are shared by design.
doc = mem.export_json("./agent_dump.json")
print(f"Exported {doc['counts']['episodes']} episodes, {doc['counts']['facts']} facts")

# Import into another store
with Engram(path="./new_store.engram") as dst:
    counts = dst.import_json("./agent_dump.json")
    # merge=True skips duplicate ids instead of raising
    counts = dst.import_json("./agent_dump.json", merge=True)

GDPR right-to-be-forgotten

# Permanently erase a single episode
mem.forget(episode_id)

# Permanently erase a single semantic fact
mem.forget_fact(fact_id)

# Erase everything about a person: episodes, facts, graph edges
result = mem.forget_entity("Ivan")
print(f"Deleted {result.episodes_deleted} episodes, {result.facts_deleted} facts")

CLI

Engram ships a command-line interface for inspecting and operating stores without writing code:

engram inspect     <path> [--agent-id ID]
engram recall      <path> <query> [--k K] [--mode cosine|spreading|hybrid] [--as-of DATE]
                                  [--agent-id ID] [--cross-agent]
engram timeline    <path> <entity>
engram observe     <path> <content> [--actors NAME...] [--tags TAG...]
                                    [--salience F] [--valence F] [--agent-id ID]
engram reflect     <path> [--llm anthropic|openai] [--model MODEL]
                          [--base-url URL] [--agent-id ID]
engram forget      <path> (--episode ID | --entity NAME) [--agent-id ID]
engram list-agents <path>

# Inspect a store
engram inspect ./agent.engram

# Store: ./agent.engram  (1.4 MB)
#   Episodes:       1842   (vec index: 1842)
#   Facts:           234   (active: 198, superseded: 36)
#   Entities:         41
#   Reflections:      12   (last: 2025-05-11 09:14 UTC)

# Recall (cosine, hybrid, or spreading)
engram recall ./agent.engram "Ivan employer" --k 3
engram recall ./agent.engram "Ivan employer" --mode hybrid --k 5

# Recall as of a past date
engram recall ./agent.engram "Ivan employer" --as-of 2024-03-01

# Observe from the command line
engram observe ./agent.engram "Alice promoted to VP Engineering" --actors Alice --tags hr

# Run reflection
engram reflect ./agent.engram --llm anthropic --model claude-haiku-4-5-20251001

# Forget an entity (GDPR)
engram forget ./agent.engram --entity Ivan

# Multi-agent: list all agents
engram list-agents ./team.engram

# Recall scoped to one agent
engram recall ./team.engram "migration" --agent-id agent://acme.example/coder

API reference

Every method on Engram and AsyncEngram, all six LLM adapters, the LangChain / LlamaIndex / MCP integrations, the on-disk schema, and the benchmark numbers live in a dedicated reference doc so this README stays readable. Start there for exact signatures, return types, and copy-paste examples for observe, recall, assert_fact, reflect, timeline, why, forget_entity, compress, WorkingMemory, and everything else.

See the full API reference.

Development

git clone https://github.com/taipanbox/engram
cd engram
python -m venv .venv && source .venv/bin/activate
pip install -e '.[dev]'

pytest -x           # run tests, stop on first failure
ruff check . --fix  # lint + auto-fix
ruff format .       # format
mypy engram         # type check (strict)

Test coverage (520 tests, 527 with the encryption extra)

Statement coverage: 89% (2679 statements, 293 uncovered on a plain install), measured 2026-09-16 with pytest --cov=engram --cov-report=term-missing. CI's test job runs the suite with the same flags and prints the current figure on every push; this line is not machine-checked against that output (unlike the test count above), so treat CI's own number as current if the two ever disagree. No threshold is enforced; the intent is a visible number, not a build that fails under one.

tests/
  test_schema.py         schema + SQLite migrations (incl. backward compat)
  test_observe.py        observe() + embeddings
  test_recall.py         cosine recall
  test_hybrid_recall.py  hybrid BM25 + cosine recall, FTS index population
  test_smoke.py          end-to-end Engram class
  test_importance.py     decay formula
  test_decay.py          decay background job + access log
  test_store_facts.py    fact CRUD + assert_fact()
  test_reflection.py     reflection loop (stub LLM), cost_tokens, reflect_async
  test_graph.py          entity/edge CRUD + spreading recall
  test_bitemporal.py     as_of + timeline (incl. naive-datetime boundaries)
  test_forget.py         forget(), forget_entity(), GDPR cascade
  test_cli.py            all CLI subcommands + --agent-id + --cross-agent
  test_multiagent.py     agent_id scoping, shared facts, cross-agent recall
  test_performance.py    observe_many correctness + batch decay + LRU cache
  test_export.py         export_json / import_json round-trip, merge mode,
                          agent-scoped export
  test_backup.py         backup() - hot copy, openable as Engram, refused on a
                          scoped instance
  test_working_memory.py WorkingMemory LRU, eviction, flush, spillover
  test_async_engram.py   AsyncEngram - all async methods + concurrency safety
  test_compress.py       compress() - LLM summarisation, batching, no-op paths
  test_encryption.py     SQLCipher encryption-at-rest + rekey()
  test_llm_adapters.py   all LLM adapters + response-parsing edge cases
  test_integrations.py   LangChain, LlamaIndex
  test_mcp_server.py     MCP server (engram-mcp): remember/recall/why/forget/stats,
                          agent pooling, procedural rejection, reflect() not exposed
  test_events.py         Agent Passport NDJSON event exporter: schema validation,
                          fail-open on I/O error, skip-on-empty agent_id, warn-on-
                          nonconforming agent_id, off-by-default, bulk and cascade
                          paths (observe_many, forget_entity, compress), the v0.2
                          envelope version, a chain that continues across it, and
                          a delegation proof accepted but never written
  test_gates_are_wired.py  the gate scripts run in the workflows that ship
  test_benchmarks.py     benchmark infrastructure

Status

Contributing

PRs welcome. Please:

Open an issue first for non-trivial changes.
Follow Conventional Commits (feat:, fix:, refactor:).
Run pytest -x && ruff check . && mypy engram before submitting.
Keep PRs small - one logical change per PR.

See CONTRIBUTING.md for the full development guide.

License

Apache-2.0 - see LICENSE.

Architecture rationale and design decisions: DESIGN.md
Release notes by version: CHANGELOG.md
Read / write paths and on-disk guarantees: DATA_FLOW.md

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
benchmarks/results		benchmarks/results
docs		docs
engram		engram
scripts		scripts
tests		tests
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
DATA_FLOW.md		DATA_FLOW.md
DESIGN.md		DESIGN.md
GETTING_STARTED.md		GETTING_STARTED.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
VALIDATION.md		VALIDATION.md
components.json		components.json
pyproject.toml		pyproject.toml

Latest commit

History

Folders and files

Repository files navigation

Engram: Cognitive Memory Layer for AI Agents

The problem

What Engram does

Where this fits in the stack

Live infrastructure validation

Running it on a Kubernetes cluster

What is Engram, technically?

What you can actually do with it

What the memory actually holds

How Engram works

Memory that doesn't forget the wrong things

Facts that know when they were true

What a recall actually does

Three ways to search

How well it recalls, measured

A scratchpad for the agent's current task

Background reflection (the agent's "sleep")

Compressing old memories

Under the hood - technical details

Bitemporal validity

Hybrid BM25 + cosine recall

Importance scoring formula

Spreading-activation graph traversal

Working memory - Miller's 7±2 law

Install

Verify a download

Quickstart

Basic usage

Async API

Working memory scratchpad

Hybrid recall

Bulk import with observe_many

Async reflection with Claude

Memory compression

Time travel

Multi-agent shared store

Backup and export

GDPR right-to-be-forgotten

CLI

API reference

Development

Test coverage (520 tests, 527 with the encryption extra)

Status

Contributing

License

About

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages