traqr.dev·working document·sheet 02 — memory
A memory database for coding agents.
TraqrDB is the piece I built first and the piece everything else leans on. Agents store what they learn while they work — bug root causes, conventions, your preferences — and load it back before the same mistake happens twice. It’s a schema and twelve MCP tools on top of your Postgres. No service, no account; nothing leaves your database.
§ 00·the loop
Store. Search. Context.
The three calls an agent makes most. A gotcha goes in mid-session as a one-liner; weeks later a different agent asks a vague question and gets it back, ranked and cited.
fig. 1 — the shape of each call. Illustrative session; the tool output format is the real one.
§ 01·the record
Not a text dump.
Each memory is a structured record. You hand the store call one string; the rest is derived on write, so search has something better than raw prose to rank against.
content + summarywhat you stored, plus an auto-derived one-liner for cheap scanningcategory · topic · tagsderived from the content on write — gotcha, fix, preference, patternentitiesnames extracted and linked, so “Vercel” connects across memoriesconfidence0–1, defaults low on purpose — bad context is worse than noneversion chaincorrections supersede instead of overwrite; the history stays readable§ 02·the tools
Twelve tools.
Everything the server registers, verified against the source before printing. Capture, retrieval, and the hygiene tools that keep a long-lived database honest — including hard-delete for when honest means gone.
memory_storestore a learning; category, tags, topic auto-derivedmemory_enhanceextend a topic with a new connected memorymemory_pulsebatch capture + search + update in one callmemory_searchsearch by meaning; semantic + BM25 fusedmemory_readexpand one memory: content, versions, relationsmemory_contextassemble principles, preferences, gotchas for a taskmemory_browsenavigate by facet; zero embedding costmemory_auditsystem health, stats, quality metricsmemory_archivehide stale content that was once correctmemory_correctstore the fix, archive the wrong one, link themmemory_forgethard-delete one memory; audit trail keptmemory_purgeright-to-delete a whole namespace, export first§ 03·your infrastructure
One SQL file, your database.
Setup is setup.sql applied to a Postgres you already run — Supabase, RDS, Cloud SQL, a Docker container. Embeddings come from OpenAI, Gemini, Bedrock, or Ollama running locally; or skip embeddings entirely and run BM25-only.
fig. 2 — what setup actually creates, counted from setup.sql on 2026-06-10.
§ 04·install
Start here.
The memory server stands alone — most people start with it. It’s also the first thing the full harness installs.