local-first · MCP stdio · 8 tools

Evidence memory for coding agents

Zephr is the evidence and currentness layer underneath AI coding workflows. Recover what is known, why it is believed, when it was true, and what contradicts it — across context window collapses and tool switches.

The four questions

A recalled belief must answer four questions

Zephr prefers abstention over fluency when evidence is insufficient. Inspectable memory that survives a context window collapse and travels between agents without losing provenance.

01
Where did this come from?

Session, commit, and file-line provenance for every belief.

02
When did it apply?

Bi-temporal recording — know if context is current or stale.

03
What contradicts it?

Structured contradictions with review state, not silent conflicts.

04
Why was it retrieved?

Full score decomposition: lexical, semantic, and RRF fusion.

What makes it different

Trust over fluency

Not another note store. Inspectable memory with deterministic recall, provenance anchors, and structured abstention.

Provenance chains

Every recalled belief carries a full chain: session → commit → file:line. No mystery sources, no hallucinated context.

Bi-temporal evidence

Beliefs track when they were recorded and when they were true. Stale context is flagged, not silently served.

Contradiction detection

When two beliefs conflict, Zephr surfaces the contradiction with both sides — it abstains rather than guessing.

Hybrid retrieval

FTS5 lexical + sqlite-vec semantic, fused via reciprocal rank. Score decomposition is inspectable per result.

Local-first

Your evidence graph lives in a SQLite file on your machine. No cloud, no telemetry, no account required.

8 tools, forever

A frozen MCP surface: remember, recall, why, verify, session, rules, status, admin. No ninth tool. No drift.

Start in 30 seconds

Add Zephr to your Claude Code or Cursor MCP config. No account, no cloud.

Read the setup guide →