Deja Vu

Deja Vu is a memory layer for coding agents built out of the session transcripts they already write to disk. It parses the JSON-lines and SQLite session stores of 22 harnesses in place, strips credentials, and builds a local inverted index, so history from before it was installed is searchable — including sessions whose original files a harness has since cleaned up. Nothing is copied to a service, and no model or embedding key is required.

Recall reaches an agent through an MCP server exposing one tool with several modes: matching snippets for a query, a Markdown digest of the best-matching session, what was discussed and decided about a given file, what this machine ran after the same error before, the real invocation of a command as it was actually used here, and a note stored deliberately for later. Hooks supply recall without being asked — at session start, before a file is edited or a command runs, and after a command fails.

The index captures the work as well as the conversation: which files each turn opened, which commands ran and with what exit status, and the exact spans an edit replaced. A found session can be reopened in its own harness, distilled into a curated note with provenance and tags, or marked as tried and rejected with a reason so later hits carry that outcome. A hit reports how many of that session's files have changed since, and says nothing when it cannot tell. Memory moves between machines over ssh, append-only, and a live context can be packaged for handover to a different harness.

Indexing and search stay local; the network is used only for updates, ssh sync and a version check. Keys, bearer tokens, JSON web tokens, private-key blocks, credentials embedded in URLs and high-entropy values are replaced with a redaction marker at index time while the surrounding text stays searchable, and exports re-apply redaction. Semantic recall is optional: pointing the tool at a local Ollama or LM Studio runtime, or any OpenAI-compatible embeddings endpoint, adds vector matching for rephrased queries, and lexical search is unchanged without one.

Measurement	Result
LongMemEval-S	85.3% hit@1
LoCoMo	69.6%
Lookup, in process	~0.4 ms median
Query, end to end	~0.2 s median over 1,551 sessions and 143k messages (5.2 GB)
Index size	160 MB, about 3% of the corpus

Features

• Retroactive search: gigabytes of prior sessions searchable, including everything recorded before installation
• Cross-agent recall: one index answers in whichever harness asks, whoever solved the problem originally
• Automatic recall: session-start, pre-edit, pre-command and post-failure hooks surface prior context unprompted
• Action-level indexing: files opened, commands with exit status, and the exact text an edit replaced
• Decision lifecycle: sessions distilled into curated notes, markable as accepted or rejected with a reason
• Staleness signal: a hit reports how many of that session's files have changed since
• Redaction at index time: keys, tokens, JSON web tokens and private-key blocks replaced before anything is stored
• Sync and handoff: append-only transfer between machines over ssh, and packaging a live context for another harness
• Optional embeddings: vector recall through a local runtime or an OpenAI-compatible endpoint, with lexical search as the default
• 22 harnesses: per-harness support for recall, auto-recall, skills, commands, resume and handover is documented as a matrix
• Benchmarks in-repo: ranking, context and answer-carrying benchmarks run on public datasets
• Local reporting: the whole memory browsable as one local page, plus usage statistics as a terminal card or timeline