engram-memory

Engram Memory

Engram gives you memory that survives the end of this conversation. Treat it like a colleague's notebook: read it before you start, write in it when you learn something the next session will need. A session that starts with recall feels like continuity; a session that ends with a write-back makes the next one smarter.

The rhythm

Session start — recall before acting:

engram_recall { query: "<the project or topic you're about to work on>", limit: 5 }

Scope with namespace when you know the project (see Namespaces). If you're about to do substantial work in a known context, prefer engram_context — it returns a pre-formatted block of the most relevant memories in one call:

engram_context { query: "<topic>", namespace: "<project>", max_tokens: 500 }

During the session — write when durable knowledge appears (see "What deserves a memory").

Session end — after completing significant work, store the outcome: what shipped, what broke, what was decided and why. This is the single highest-value write; it's what makes the next session start from "here's where we left off" instead of zero.

What deserves a memory

Store it when a future session would act differently for knowing it:

• The user tells you to — "remember", "don't forget", "for future reference". Always store, confidence 1.0.
• Decisions with their why — "we chose X over Y because Z". The rationale is the valuable part; a decision without its why gets relitigated.
• Corrections — the user corrects a wrong assumption you made. These are gold: they prevent the same mistake in every future session.
• Preferences — how the user likes to work, tools they favor, styles they reject.
• Setup and infrastructure facts — versions, hosts, ports, conventions, account structures, "the staging DB is the one named prod2" landmines.
• Outcomes — what a completed piece of work produced, including failures and dead ends ("tried X, doesn't work because Y" saves the next session from repeating it).