recall

You solved this before. Six weeks ago, in a session you can't find.

Claude Code and Codex write every session to disk as JSONL and then give you no way to search it. recall indexes both tools into one local SQLite store and answers natural-language or keyword queries from the terminal — hybrid semantic + full-text ranking, single Go binary, no cloud service in the retrieval path.

$ recall "how did I fix the pyexpat linking" --pretty
DATETIME          SOURCE  SESSION   PROJECT     TITLE
2026-04-17 22:31  claude  4a4b4bfe  build-tools  Static-linking pyexpat into the wheel

It also doubles as a session manager: list, pin, summarize, and resume past sessions (see Command reference).

Privacy, stated precisely: indexing and search are local — but recall summary sends transcript text to the codex CLI, and a remote OLLAMA_HOST sends text off-box. Read Privacy & network behavior before you point this at sensitive transcripts.

Documentation map

Doc	What it covers
DESIGN.md	Architecture, schema, every design decision + the alternatives rejected (why SQLite over LanceDB/DuckDB/turbopuffer, why Go, why brute-force vectors).
docs/embedding-models.md	Which embedding model recall uses and why — the bge-m3 vs qwen3 comparison, results, and the decision.
docs/evaluating-embeddings-101.md	Beginner's guide — what embeddings are, how search is benchmarked, what Recall@k / MRR mean. Read this if the comparison doc feels jargon-heavy.
eval/	The benchmark that picked the default model: query set (fixture.json), harness (run.py), results (results.json). Not reproducible by outsiders — see Evaluation.

Prerequisites

• Go — only if you build from source. go.mod declares 1.26.3, so that's the minimum the toolchain will accept. The result is a single binary (no cgo). Not needed if you install with Homebrew.
• Ollama, running, with an embedding model pulled:

  ollama pull qwen3-embedding:0.6b   # 639 MB

  Required for recall index and for semantic/hybrid search. Lexical search and --all need no model.
• codex CLI — only for recall summary, which shells out to it. Must be installed and authenticated. Every other command works without it.

Developed and used on macOS (arm64). Nothing is knowingly platform-specific except the transcript locations and the resume commands, but Linux/Windows are untested.

Install

Homebrew (easiest — prebuilt binary, no Go toolchain needed):

brew install ramsrib/tap/recall

go install (installs the binary as recall-cli; rename or alias it to recall):

go install github.com/ramsrib/recall-cli@latest

From source:

git clone https://github.com/ramsrib/recall-cli
cd recall-cli
make install        # builds and installs recall → ~/.local/bin

make install warns if ~/.local/bin isn't on your PATH and prints how to add it. Install elsewhere with make install BINDIR=/usr/local/bin.

make            # list all targets
make build      # build ./recall in the repo dir only
make uninstall  # remove the installed binary
make clean      # remove the local build artifact

Or plain Go: go build -o recall .

Quickstart

ollama serve &                 # if it isn't already running
ollama pull qwen3-embedding:0.6b

recall index                   # first run embeds everything — minutes, not seconds
recall "that vector db tradeoff discussion" --pretty

recall index is incremental: re-run it anytime and it only touches sessions whose files changed. The first run is the slow one — it embeds your entire history.

Command reference

Every command prints compact single-line JSON to stdout by default (agent-first: no flag needed when a script or agent calls it). --pretty switches to a human-readable table. Progress, warnings, and notes go to stderr, so stdout stays clean JSON.

recall index [--model NAME]

Reconciles the index with disk: new and changed session files are parsed, chunked (~512 tokens), embedded via Ollama, and written to ~/.config/recall/index.db. Files deleted from disk are dropped from the index.

• --model NAME — embed with a specific Ollama model. Re-embeds only the files not already on that model, so switching models is incremental and resumable, not one big reindex. Default: $RECALL_EMBED_MODEL, else qwen3-embedding:0.6b.
• Requires Ollama to be running with the model pulled.
• Also refreshes the sessions registry that recall list reads.

recall "QUERY" [flags]

The default command — anything that isn't a known subcommand is treated as a query. Results are collapsed to one row per session (best-scoring chunk wins).

Flag	Default	Meaning
--mode hybrid|semantic|lexical	hybrid	semantic = embeddings only; lexical = FTS5/BM25 keywords only; hybrid = both, merged with Reciprocal Rank Fusion.
--all	off	Session-level AND: which session mentions every one of these words somewhere (even in different messages)? Pure lexical, no embeddings.
--source claude|codex	both	Restrict to one tool.
--project SUBSTR	—	Substring match on the session's working directory.
--since YYYY-MM-DD	—	Only sessions on/after this date.
--limit N	20	Max results.
--model NAME	the index's current model	Rank against another model's vectors (they must already be indexed under it).
--include-exec	off	Include non-interactive codex exec / headless claude -p runs, which are hidden by default.
--pretty	off	Human table (DATETIME · SOURCE · SESSION · PROJECT · TITLE) instead of JSON.

Flags may appear before or after the query. JSON fields: source, session_id, project, title, role, ts, path, score, snippet — path is the transcript on disk, so an agent can open the matched session.

recall "how did I fix the pyexpat linking"
recall --mode semantic "that vector db tradeoff discussion"
recall --mode lexical  "pyexpat"
recall "webhook retry backoff timeout" --all
recall "deploy script" --project backend --since 2026-05-01 --limit 10 --pretty

--all vs the modes. The modes match within a single chunk (one message span) and rank by relevance. --all intersects per-keyword session sets — use it to pin down one specific conversation from a handful of distinctive words that were never in the same sentence.

recall list [flags]

One row per session, newest activity first, with pin/summary state. Reads the sessions registry (not the chunk index) and refreshes it from disk first, so new sessions appear without a reindex.

Flag	Default	Meaning
--mode interactive|exec|all	interactive	Which session kinds to include. Default hides non-interactive codex exec/headless runs.
--source, --project, --since	—	Same as search.
--limit N	0 (all)	Max sessions.
--pretty	off	Table: DATETIME · SOURCE · SESSION · PROJECT · TITLE · FLAGS. Flags are 📌 pinned, ✶ has summary, ⚙ exec.

recall pin <session-id> / recall unpin <session-id>

Mark a session as pinned (surfaced first in the pinned group of a manager UI). --source claude|codex is auto-detected from the index if omitted. Purely local bookkeeping in the pins table.

recall summary <session-id> [flags]

Generates a concise overview of a session — gist, key points, outcome, next steps — and caches it in the DB, keyed on the transcript's mtime (so it regenerates when the session grows).

⚠️ This command sends transcript content off your machine. It renders the session to plain text, caps it at 48,000 runes (head + tail, middle elided), and pipes it into codex exec on stdin. Codex is normally network-backed by OpenAI, so the transcript goes wherever your Codex configuration sends it, and it costs whatever your Codex usage costs. This is the only command that does this.

Flag	Meaning
--refresh	Ignore the cache and regenerate.
--cached-only	Return the cached summary if one exists (ignoring staleness); never generate. Empty summary if none cached — so glancing at a session never spends a codex call.
--model NAME	Codex model to use. Default: whatever codex is configured with.
--pretty	Print just the summary text instead of JSON.

Requires an installed and authenticated codex CLI. Looked up on PATH and at /opt/homebrew/bin/codex, ~/.local/bin/codex, /usr/local/bin/codex.

recall resume <session-id>

Prints the command that would resume a session, plus its working directory, as JSON: {source, session_id, cwd, command, path}.

⚠️ The emitted command disables the agent's safety rails. It is claude --dangerously-skip-permissions --resume <id> for Claude sessions and codex --dangerously-bypass-approvals-and-sandbox resume <id> for Codex sessions — flags that let the resumed agent read, write, and execute without prompting you.

recall resume only prints this string. It never executes it. Nothing runs unless you copy the command and run it yourself. If you'd rather resume with approvals intact, drop the dangerous flag and run claude --resume <id> / codex resume <id> by hand. If you pipe recall resume into a shell, you own that decision.

recall log [--limit N] [--pretty]

Recent searches, newest first. Every search is logged, non-optionally, to the queries table in the local DB: timestamp, query text, mode, filters, model, result count, top sessions returned, latency. Written best-effort — a logging hiccup warns on stderr but never fails your search.

It exists to power real-usage eval sets (vs. synthetic queries), query analytics, and search history. Clear it with:

sqlite3 ~/.config/recall/index.db 'DELETE FROM queries'

Privacy & network behavior

recall is local-first, not local-only. Precisely:

Stays on your machine (with a default, local Ollama):

• Reading and parsing transcripts.
• Chunking, embedding (Ollama at http://localhost:11434), and storing vectors.
• All search — semantic, lexical, hybrid, --all.
• The query log, pins, and the sessions registry.

Leaves your machine:

• recall summary pipes up to 48,000 runes of transcript into codex exec (sessions_cmd.go). Codex is network-backed by OpenAI in normal use. If a session's content shouldn't reach a third-party API, don't summarize it. No other command calls out to anything.
• A non-local OLLAMA_HOST. It defaults to http://localhost:11434, but it's a plain env var — point it at a remote Ollama and every chunk you index and every query you type is sent to that server (internal/embed/ollama.go).

What the database actually holds. ~/.config/recall/index.db is not an opaque vector blob. It stores verbatim copies of your transcript text (the chunks table, mirrored into an FTS5 index), file paths, your full query history, pins, and generated summaries. Treat it as being as sensitive as the transcripts themselves. Its directory is created 0755 — on a shared machine, tighten it (chmod 700 ~/.config/recall).

To delete everything recall created:

rm -rf ~/.config/recall     # index, query log, pins, summaries — all of it

Deleting the index never touches your original transcripts in ~/.claude / ~/.codex.

What is and isn't indexed

Sources. Fixed, built-in locations only — there is no config for custom transcript roots, and XDG_CONFIG_HOME is not honored:

• ~/.claude/projects/<encoded-cwd>/<session-uuid>.jsonl
• ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl and ~/.codex/archived_sessions/rollout-*.jsonl (archived Codex sessions are indexed)

Indexed:

• Claude: user and assistant messages — text blocks and thinking blocks.
• Codex: user and assistant messages.

Not indexed:

• Tool calls and tool results. tool_use / tool_result blocks are explicitly skipped. Search finds the conversation, not the command output. (This is why a query for a filename you only ever saw in a tool result may come up empty.)
• Codex developer / system records.
• Claude sub-agent transcripts (<session-id>/subagents/agent-*.jsonl) — they carry the parent's session id and would create duplicate rows.
• Codex logs_2.sqlite (TRACE diagnostics) and legacy pre-2025 .json sessions.

Indexed but hidden by default: non-interactive runs (codex exec, claude -p/SDK). Pass --include-exec to search them or --mode exec|all to list them.

Config

Env	Default	Purpose
RECALL_EMBED_MODEL	qwen3-embedding:0.6b	Embedding model at index time. See the caveat below.
OLLAMA_HOST	http://localhost:11434	Ollama server. Non-local values send your content off-box.

⚠️ RECALL_EMBED_MODEL does not control search. At search time (without an explicit --model), recall reads the model from the DB's own metadata — the model the index was last built with — and ignores the env var. This is deliberate: a query must be embedded with the same model as the vectors it's compared against, or cosine similarity is meaningless. Precedence for search is --model → the index's recorded model. Precedence for indexing is --model → $RECALL_EMBED_MODEL → qwen3-embedding:0.6b.

Index lives at ~/.config/recall/index.db.

Switching embedding models

Each file records the model that embedded it. recall index --model NAME re-embeds only the files not already on that model — incrementally and resumably, not one big reindex — so you can A/B models cheaply, and an interrupted migration just leaves a mix that the next run finishes.

Semantic search only ranks chunks embedded with the query model; sessions still on another model are skipped from semantic ranking (with a note on stderr) until you re-embed them. Lexical and --all are model-agnostic and always see every chunk.

The default is qwen3-embedding:0.6b, chosen by a head-to-head eval against bge-m3 (both 1024-dim, so switching is drop-in) — see docs/embedding-models.md.

Evaluation (and why you can't reproduce it)

eval/ holds the labeled benchmark that picked the default model: 22 natural-language queries, a harness that drives the real recall binary, and the resulting Recall@k / MRR scores.

It is not reproducible by anyone but the original author, by construction. The gold answers in eval/fixture.json are session-id prefixes from a private local transcript corpus. You do not have those sessions, so on your machine every query scores zero. The harness and the methodology are the reusable parts — the numbers are not independently verifiable, and you should read them as a documented decision rather than a public benchmark. To evaluate on your own corpus, replace fixture.json with queries and gold session ids drawn from your index.

Troubleshooting

ollama unreachable at http://localhost:11434 (is it running?) Start it: ollama serve. If Ollama lives elsewhere, set OLLAMA_HOST — and note that a remote host means your content leaves the machine.

ollama /api/embed 404: model "qwen3-embedding:0.6b" not found ollama pull qwen3-embedding:0.6b. Confirm with ollama list.

index is empty — run 'recall index' first Semantic/hybrid search needs a model recorded in the DB, which only recall index writes. If recall index reported 0 sessions/files, nothing was discovered — check that ~/.claude/projects or ~/.codex/sessions actually exist and contain .jsonl files.

Search returns nothing, but you know the session exists. In order of likelihood: (1) it's a codex exec/headless run — add --include-exec; (2) the text you're searching for only ever appeared in a tool result, which isn't indexed; (3) the session was indexed under a different embedding model — check for the stderr note and re-run recall index; (4) --project / --since filtered it out.

note: N session(s) embedded under another model were excluded from semantic ranking Your index is mixed. Run recall index to bring everything onto the current model, or query the other model explicitly with --model.

First index is very slow. Expected — it embeds your entire history (roughly 5 chunks/s for qwen3 on an M-series Mac). It's one-time; later runs only touch changed files.

codex CLI not found (needed for summaries) Only recall summary needs it. Install and authenticate the codex CLI, or don't use that command.

Corrupt or incompatible index. There's no repair path and no version negotiation — the index is a derived artifact. Delete it and rebuild: rm -rf ~/.config/recall && recall index. Schema upgrades that can be done in place (adding the per-file model columns) happen automatically on open.

Roadmap

• Reranker stage (bge-reranker-v2-m3) — the Reranker extension point is already wired in internal/search.
• --full to print the matching transcript span.
• Tests. There are none today; that's a gap, not a decision.

Related

• recall-app — desktop session manager built on this CLI. list, pin, summary --cached-only, and resume exist to serve it, and are equally usable on their own.

License

MIT — see LICENSE.