This guide walks you through the core Engram workflow end to end: installing the tool, staging a low-risk fact and a sensitive one, running the sync bridge to promote or queue them, reviewing and approving the queued fact, recalling your memory registry, and finally wiring Engram into a coding agent so every new session starts with full context.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/xantorres/engram/llms.txt
Use this file to discover all available pages before exploring further.
Engram is not yet on PyPI, so install it directly from GitHub. Pick whichever tool you already have:
Use
engram remember to stage facts. Engram doesn’t write them to the registry immediately — they go into a staging area first so you can control what gets promoted.The
tooling kind is tier 1 (auto-eligible). The fiscal kind is tier 3 and will always be held for human review regardless of any environment variable.Set
ENGRAM_AUTOPROMOTE=true to let tier-1 facts promote automatically, then run engram sync with --apply to execute (omitting --apply runs a dry-run):# append mem-0001 [tooling] I prefer pnpm over npm (low-risk)
# queue mem-0002 [fiscal] My VAT number is ... (fiscal needs review)
The tooling fact was appended to
memory.md immediately. The fiscal fact was moved to queue/ and is waiting for your approval.Skip
export ENGRAM_AUTOPROMOTE=true if you prefer to review every fact manually before promotion, regardless of tier.You’ll see the raw JSON including the kind, confidence score, timestamp, and the original fact text — everything you need to make an informed decision.
Once you’ve reviewed
mem-0002 and are satisfied it’s correct, promote it with the required --confirm flag:The
--confirm flag is mandatory for tier-3 kinds (fiscal, identity, people, constraint, location, health). Engram will refuse to promote without it as a guard against accidental approvals.Then generate a curated context block and write it to
AGENTS.md so agents pick it up as static context even without MCP:Start the Engram MCP server over stdio so your agent can call
recall, remember, and other tools live during a session:Harvesting From Existing Sessions
If you have existing session transcripts (for example Codex JSONL logs), you can mine them for facts in bulk rather than entering each one manually:What to Do Next
- Run
engram doctorto surface stale, low-confidence, unverified, or conflicting memories in your registry. - Use
engram list --status pendingto see everything still in staging. - Use
engram forget <memory-id>to retract a promoted memory — Engram emits an undo token so the operation is reversible. - Read the Memory Model docs to understand kinds, tiers, confidence scores, and status transitions in depth.