Total Recall

Every session you run is already on disk as append-only JSONL: every decision, every "no, do it this way", every dead end. Total Recall reads that history locally and feeds the high-signal parts back to new sessions in a low-token form, so the model stops re-asking what you already told it.

Install

Marketplace (recommended):

/plugin marketplace add 88plug/claude-code-plugins
/plugin install total-recall@88plug

Local checkout (for development):

git clone https://github.com/88plug/total-recall.git
cd total-recall
pip install -e .[vec]
claude --plugin-dir "$PWD"

[!NOTE] Requirements are bash + curl + internet. The plugin bootstraps everything else (uv, Python, deps) into its own data dir on first hook fire. No system-wide pip install and no system Python required.

Quickstart

First run backfills your existing transcripts in the background (detached, so it survives the spawning session exiting; progress goes to logs/bootstrap.log). After that, every new session gets a short SessionStart brief about what is relevant to the current directory.

Check what was indexed:

/recall-status

Ask the model to use its memory at any time:

/recall what did we decide about the deploy pipeline?

For a manual full reindex:

total-recall index --rebuild --jobs 4

A typical corpus drops from about 22s single-threaded to about 9s at --jobs 4.

Why this exists

Claude Code has three kinds of memory today, and none of them mine the transcript history: amnesia (88plug) keeps one session alive across compaction but ignores other sessions; auto-memory (~/.claude/projects/<proj-slug>/memory/) is hand-curated; CLAUDE.md is static and hand-edited.

The operator is the source of truth. Models change and projects come and go; the human running the sessions is the one constant. Total Recall makes that explicit: an operator profile and voice profile are first-class extracted artifacts, queryable in one MCP call at session start.

What it captures

17 extractors total. 11 run inline over each session's record stream; 6 are operator-level aggregators that run out-of-band against the full corpus.

Per-session extractors (11)

• corrections — turns where you redirected the model.
• decisions — "we're going with X because Y" moments.
• self_corrections — places the model corrected itself ("actually, scratch that").
• progress — how far a line of work actually got; anchors "we already did X".
• domain_facts — durable signals about the codebase or environment (versions, paths, conventions).
• away_summaries — recap text you wrote after returning to a stale session.
• model_corrections — corrections specifically about model behavior or output format.
• standing_decisions — decisions you marked as durable across sessions.
• bans — explicit "never do X" instructions.
• goals — what you said you are trying to achieve in a session.
• truth_rhetoric — assertions you made about objective state, kept so a later session can check whether they still hold.

Operator-level extractors (6)

• operator_profile — durable signals about the human: who they are, how they work, preferences across projects.
• voice_profile — how you write: tone, phrasing patterns, verbal tics, so a model can match register without being told.
• ontology — vocabulary you use for your own systems (project, machine, and service names) plus a cross-project co-mention graph.
• workflow — how you work: fan-out vocabulary, autonomy score, mid-flight interrupt rate, planning idiom, preferred work window, subagent adoption.
• implicit_preferences — preferences expressed by behavior rather than as a ban or decision (tool-call ratios, shell-command dominance, format preferences). Promoted only past a multi-axis threshold.
• satisfaction — bidirectional praise/frustration profile paired with the preceding assistant-turn shape.

Reference

MCP tools (26)

Live queries the model can call mid-conversation.