Recall

/recall: session memory engine

Claude Code and Codex sessions on this machine are indexed into a local SQLite engine. Query the index and read only the best-supported session instead of searching entire transcripts. All commands go through one CLI:

python3 scripts/recall.py        # relative to this skill directory

If RECALL_URL is set, RECALL_MODE is remote or shadow, or ~/.config/recall-brain/client.json exists, this machine has a central Recall Brain configured. In that case, read [references/central-brain.md](references/central-brain.md) before running commands. Otherwise everything below is fully local and nothing touches a network.

No index yet? Search anyway

If search reports the index does not exist (or doctor shows db exists=False), search the raw JSONL transcripts immediately while the first index builds:

rg -l -i "" ~/.claude/projects ~/.codex/sessions   # candidate files
ls -t                                               # newest first
rg -n -i -C3 ""                          # read the window

Choose terms and regular expressions based on the request. Exact identifiers are stronger evidence than general prose. If rg is unavailable, use grep -rl. Start the index in the background at the same time:

setsid nohup python3 scripts/recall.py index >/dev/null 2>&1 &

The first build over a large history can take many minutes; later runs are incremental and fast. Tell the user when an answer came from a cold scan of the raw transcripts. Once doctor shows a healthy db, switch to indexed search, which ranks results, explains the WHY for each match, and matches identifiers exactly.

First: pick the outcome

1. Find / verify: answer "did we…", "which session…", "how did we…".

Search, read the best hit's relevant window, answer with the session path as the receipt.

1. Continue: resume in-progress work. This needs the session's tail plus its

branch and worktree.

1. Repeat: redo the same kind of task with fresh inputs. This needs the

original driving prompts, verbatim.

1. Skill-ify: turn the recipe into a reusable skill. This needs the steps that

worked, minus one-off data. Chain into the harness's skill creator when one is installed; otherwise write the standard SKILL.md package directly.

Ask only if the outcome is genuinely ambiguous.

Search

python3 scripts/recall.py search "" [filters]

• Pass the user's phrasing plus any identifier you have. Identifiers (job

UUIDs, PR numbers, pod names, error strings, filenames) are the strongest evidence and are matched exactly, including inside tool output.

• Apply filters as flags rather than approximating them in the query text:

| ask | flag | |---|---| | "last 48h", "back in May" | --since 2026-05-01 --until 2026-06-01 (UTC; both bounds include the specified instant. To cover a full local day, use the NEXT day's date as --until; convert the user's local day first) | | "in the other worktree/checkout" | --cwd | | "what did codex do" | --harness codex | | branch-scoped | --branch |

• Output is ranked sessions with date, cwd, slot, branch, a matched snippet,

and WHY it matched. Empty output means nothing cleared the evidence gate; it does not prove the work never happened. Retry once with a distinctive identifier or a wider window. If output is still empty, tell the user what you searched and that no supported match was found.

• --paths prints bare file paths (for scripting); --limit N widens.

Read the best-supported hit

Check the WHY line on the top few results first. Matches on only generic words are weak evidence. Prefer a result whose WHY includes an identifier, phrase, or exact-entity match rather than selecting rank 1 automatically.

python3 scripts/recall.py show  --prompts                    # user prompts only
python3 scripts/recall.py show  --around 2026-07-03T14:20    # ±3-turn window; use the date printed in the search result
python3 scripts/recall.py show  --tail 30                    # the session's final turns (for Continue)

Pass the session file path from the search output. Use show rather than cat: sessions can reach 80 MB, and show parses and prints only the requested content.

Related work (no query needed)

python3 scripts/recall.py related --cwd "$(pwd)" --branch "$(git branch --show-current)"

This returns sessions that share the project, branch, or touched files, ranked by overlap and recency. Use it at session start when the user references prior work without naming it.

Outcome playbooks

Find / verify: search → show --around the matched timestamp → answer with evidence. This usually requires two commands.

Continue: search → show --tail 30 for the final state (last actions, tool results, open errors) → check the session's branch/slot still exists (git -C branch --show-current) → summarize: "Found ` in on , last action , branch `; resume there or here?"

Repeat: search → show --prompts → present the driving prompts verbatim and confirm fresh inputs (dates, scope) before re-running.

Skill-ify: search → show the working window → separate the durable recipe (commands, endpoints, auth patterns) from one-off data (specific IDs, dates) → invoke the available skill creator with the recipe and a proposed name, or create a standard Agent Skills directory when none is installed.

Export one exact session for another skill

Use the machine-readable session export when /recap or another evidence consumer needs complete, ordered coverage rather than a human window:

python3 scripts/recall.py session-export --current --limit 1000
python3 scripts/recall.py session-export --target  --limit 1000
python3 scripts/recall.py session-export --cursor  --limit 1000

Each JSON page contains stable evidence IDs, redacted text and digests, sanitized typed entities (including native tool identity when observed), native session identity, projection/privacy versions, a boundary receipt, a content-free page receipt, and complete plus next_cursor. Consume pages in sequence and accept immutable-snapshot completeness only on the final page; inspect source_snapshot_stable before claiming a live source did not advance. Cursors are stored owner-private under ~/.recall and never encode transcript text or a path.

--current resolves Codex only through exact CODEX_THREAD_ID, and Claude through exact CLAUDE_SESSION_ID when the harness exposes it. Otherwise it fails closed with content-free ranked candidate receipts; pass the exact path found by Recall rather than guessing. Child and continuation sessions are separate boundaries by default. A standalone local export stamps an explicit local: source.

To resolve a local native relationship graph for Recap without reading transcript prose, use:

python3 scripts/recall.py session-relations --current --include-children
python3 scripts/recall.py session-relations --target  --chain
python3 scripts/recall.py session-relations --target  --chain --include-children

The closed recall.session-relations.v1 JSON uses Claude sessionId/agentId sidechain metadata and Codex parent_thread_id/forked_from_id metadata. It excludes merely adjacent or similar sessions and fails when a requested native link is missing or ambiguous. This command is local-only.

Index health

python3 scripts/recall.py index      # incremental; run if results look stale
python3 scripts/recall.py doctor     # coverage, index age, retention watchdog

Never run index --rebuild without the user's explicit request; large session histories can require gigabytes of temporary WAL and substantial CPU time.

doctor warning about cleanupPeriodDays means transcript retention got re-enabled. Surface that to the user immediately because history is being deleted.

Gotchas

• The engine indexes user text, assistant text, and tool input/output, but

reasoning/thinking blocks are never stored, and secret-shaped lines are redacted at ingest. If the only trace of something was a thinking block, it is not findable.

• Codex sessions are one file per rollout under a date tree; show handles

both schemas transparently.

• Running Recall from pi is supported, but pi's own session format is not yet

indexed. Do not claim that a cold result proves no pi session exists.

• A query about work that never happened can still return lexically-adjacent

sessions. The ranked WHY line tells you what actually matched; read it before asserting the session answers the question.

• Subagent and workflow transcripts are indexed as their own sessions and

live under the parent session's directory (/subagents/…); the path itself tells you which main session spawned them.

Upgrade: central Recall Brain (optional)

Recall can sync into a private central Brain service that provides deliberate memory writes, cross-device search, consented ChatGPT-export import, a Cowork collector, pull connectors, and MCP capture. The service is off unless explicitly configured. Setup, mode routing, and all Brain commands are in [references/central-brain.md](references/central-brain.md); doctor prints the current mode.

References

• [references/query-cookbook.md](references/query-cookbook.md): worked

examples per stratum: identifiers, error strings, time windows, cross-worktree, cross-harness, paraphrase.

• [references/central-brain.md](references/central-brain.md): optional

central Brain upgrade: setup, modes, deliberate writes, connectors, privacy, export inbox, MCP capture.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Parcha-ai
• Source: Parcha-ai/parcha-skills
• License: MIT

Install and usage instructions live in the source repository linked above.