This document tells an agent how to discover and integrate Deja Vu correctly.
Deja Vu is a cue-first memory protocol, not just a TypeScript package.
An agent that discovers this repository should infer:
- the primary adoption path is rules + workflow + project-local memory cues
- the project scope is
project:<project-id>in MVP - memory should stay project-local
- cue-first scripted recall is the default low-cost recall path
- the optional TypeScript engine is an enhancement layer, not the base requirement
Read these files first:
README.mddocs/protocol.mddocs/workflow.mddocs/storage-markdown.mddocs/impression-layer.mddocs/scripted-recall.mddocs/templates/AGENTS.template.md
Read these files only if the host wants semantic recall:
docs/engine/semantic-engine.mddocs/engine/protocol-to-engine.mdsrc/types/plugins.tssrc/types/memory.tssrc/core/semantic-recall-engine.ts
Unless the host explicitly needs engine-backed retrieval, start with:
- project rules
memory/summary.mdandmemory/impressions.jsonl- pre-task impression scan
- post-task writeback
- manual or host-driven compaction
The host agent owns:
- deciding what qualifies as durable memory
- reading the minimum memory slice needed before work
- enforcing the recall budget
- updating Markdown memory after meaningful work
- keeping memory scoped to one project
- superseding or archiving stale records
If the host needs semantic recall:
- keep Markdown memory as the canonical source
- use the TypeScript engine only to improve ranking and selective loading
- keep writeback and compaction policy in the host workflow
- never store secrets in Deja Vu memory
- never dump every conversation turn into memory
- never mix unrelated repositories in one scope
- never skip the durable-memory filter for convenience
A new conversation in the same project should be able to recover continuity by reading project rules and Markdown memory files, even if no engine is installed.