Engram — Agent Memory

Local semantic memory with biological decay, typed memories, and relationship graphs. No API keys. No cloud.

Boot Sequence

engram search "<current task or context>" --limit 10

Always recall before working. Accessed memories get salience-boosted.

Storing

engram add "Client uses React with TypeScript" --type fact --tags react,client
engram add "We decided to pause ads" --type decision --tags ads
echo "Raw conversation text" | engram ingest

Types: fact, decision, preference, event, relationship

Searching

engram search "what tech stack"
engram search "pricing decisions" --type decision
engram search "client status" --agent client-agent

Relationships

engram relate <src> <tgt> --type supports
engram auto-relate <id>
engram relations <id>

Types: related_to, supports, contradicts, caused_by, supersedes, part_of, references

Key Concepts

• Decay: Unused memories lose salience daily. Recalled ones get boosted.
• Types: Filter by fact, decision, preference, event, relationship.
• Scoping: global, agent, private, shared.
• Dedup: >92% similarity auto-merges.

Quick Reference

engram stats
engram recall --limit 10
engram export > backup.json
engram import backup.json