Recall, verbatim memory for AI coding agents.

Constraint

Three failure modes recur in AI coding tools.

Every session starts from zero. The architecture, the constraints, the decisions made yesterday all have to be re-explained at the top of every new session.

Compaction is lossy. When the context window fills, the model summarises. The summary loses the file paths, the line numbers, the specific decisions, and the rationale that produced them.

Per-tool, not per-you. What you said to Claude Code in March does not exist when you open Codex in May. Memory, where it exists, is owned by the tool, not the person using it.

First principles

Memory in AI tools fails for a reason that has nothing to do with the database. The model paraphrases. Paraphrase is lossy by construction: it preserves the gist and discards the artefacts. A gist can answer what did we decide. It cannot answer which line in which file did we cite, or what trade-off did we explicitly reject, or what was the exact PR that closed it.

Why paraphrase loses

Recent versions of Claude Code ship a built-in memory file. It is paraphrased by the model and capped at 200 lines / 25 KB. That fixes the worst case (full re-orientation every session) at the cost of fidelity.

What paraphrased memory cannot return is the verbatim quote you said three weeks ago. It cannot return the trade-off you considered and rejected. It cannot return the PR you cited when you picked Postgres over Mongo for billing. If those details matter to the next session, paraphrased memory is the wrong tool.

Mechanism

Recall keeps the verbatim text of every session you have had with any compatible tool, on your machine. Quotes come back exactly as you wrote them, with timestamps and source context. Decisions come back with the trade-offs that produced them. References (PRs, file paths, line numbers) come back with the references intact.

Architecture is local-first. Session transcripts live on disk. The index is yours, not a vendor's cache. Retrieval is hybrid: lexical for verbatim quotes, semantic for adjacent context.

Trade-off

Verbatim is bigger and slower than paraphrased. We name this rather than pretending we dominate. The trade-off is the positioning. We kept your real words. The cost is local storage and hybrid retrieval. If you do not need verbatim, the built-in summary is the right tool. If you do, Recall is.

Where it came from

Recall was built for the lab's own instances. We run five across Claude Code, Codex, and OpenClaw, and the same gap recurs at every machine boundary. The verbatim layer is what let our PA instance pull the actual decision from a daily review rather than its own paraphrase of it. We are releasing it because the same gap shows up in any multi-tool setup.