Skip to content
persistent memory for your agentsthe 0.8.1 update

engram.

good work deserves
a longer memory.

keep the decisions, context, and connections.
pick up where you left off.

start remembering read the docs
your store. your models. your context.follow the trace
01 / continuitythe next conversation starts here

a conversation ends.
the context stays.

engram keeps what matters in a store you control. decisions, notes, and procedures stay connected to their sources, ready for the next piece of work.

  1. 01

    leave a trace.

    save a decision or bring in your notes. keep where it came from, when it mattered, and the project it belongs to.

    a decisionuse sqlite for the local install.
  2. 02

    find the thread.

    ask in your own words. semantic search, keywords, and connected entities work together to find relevant context.

    a later questionwhat did we choose for storage?
  3. 03

    see why it returned.

    inspect the ranking and relevance gates. understand why a memory was returned or rejected before relying on it.

    an inspectable resultthe decision, its source, and its retrieval trail.

an illustrative workflow. a recalled memory is context, not proof that its claim is still current.

02 / measured retrieval

the right session.
in the first five.

read the evidence

470/470

100.0% session recall-any@5

a fresh LongMemEval run on the development set used during tuning. 30 abstention questions excluded; production confidence gate not applied. this measures session retrieval, not answer accuracy.

03 / clear by designbuilt to be understood

your memory.
your settings.
no guesswork.

start locally with sqlite and local models. choose postgres or hosted models when your work calls for them. keep the controls visible.

01

one configuration.

validated settings, clear defaults, and a safe view of what is actually in effect.

 02

a retrieval trail.

see the candidates, ranking signals, and reasons a memory passes or misses a gate.

 03

room to change.

edit, challenge, supersede, or forget a memory as the work moves on.

 04

meet your tools.

connect through MCP, use the CLI or local API, and inspect the store in the web workspace.

 05

build it into your app.

give NPCs a memory, keep a server handoff, or bring context into a chat. start from usable source and detailed guides.
04 / pick up the thread

make the next
conversation count.

pip install engram-memory-system

installation guide connect your agent explore the source

python 3.11+ · sqlite by default · local models download on first use