Amnesia

Install

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

[!TIP] That is the whole setup. After install, amnesia runs entirely in the background — no commands to run, no files to manage, no visible interruptions.

Quickstart (under 60 seconds)

1. Install with the two commands above.
2. Work normally. Let a /compact (or auto-compact) happen mid-task.
3. On the next turn, the agent already knows what it was doing — the working theory, the next action, the files in motion, and your most recent constraints — instead of resuming from a lossy summary.

To confirm amnesia is active, run /amnesia:status and you will see the captured handoff for the current project.

What it does

When Claude Code's /compact (or auto-compact) fires, the model's working memory is replaced by a one-shot summary. Exact file contents, prior tool outputs, verbatim user constraints, and mid-task reasoning are all paraphrased away, so long sessions drift after every compact.

amnesia makes compaction non-destructive. It captures a structured handoff at every compaction boundary — and proactively before the next one — then re-injects it on the next session start. The agent resumes with explicit knowledge of:

• what it was just doing (working theory plus the concrete next action)
• which files were in motion and which commands had run
• the user's most recent constraints, verbatim
• the path to the on-disk JSONL transcript, which Claude Code never truncates, so any lost detail is recoverable for the cost of one Read

[!NOTE] Recommended install is via the 88plug marketplace. Installing directly from this repo also works but is less convenient and will not auto-update with curated releases.

How it works

amnesia runs as background hooks. Each event is invisible to you.

Event	When	Latency	What it does
Continuous capture	Every tool call	<50 ms sync	Appends a one-line record to working-state.jsonl
L1 mechanical	After every compaction	<500 ms sync	Writes a deterministic handoff from JSONL plus working-state
L2 enrich	After every compaction	~30-60 s async	Opus 4.7 --effort max rewrites the handoff with narrative
L3 refine	First Stop after compaction	~30-60 s async	Opus refines the handoff using what happened post-compact
Preemptive	At ~75% of context	~30-60 s async	Snapshots state before the next compact, while it is still in the window
Restore	SessionStart (compact / resume / startup)	<100 ms sync	Cats the handoff into additionalContext for the next turn

All summarization runs isolated from your CLAUDE.md, auto-memory, and auto-triggered skills (via CLAUDE_CODE_DISABLE_CLAUDE_MDS=1 and CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 plus a strict system prompt). Without this, the summarizer hallucinates content from your global context — verified empirically and fixed in v0.2.0.

Cost on subscription plans

Every summarizer invocation is claude -p against your OAuth credentials. On a Max-style subscription this does not bill per token in dollars; it draws from your plan quota.

Per-call cost detail (verified 2026-05-24)

Per L2/L3/preempt call:

• ~33K Opus 4.7 cache-creation, ~21K cache-read, ~3K output tokens
• ~$0.20 informational (zero actually billed on subscription)
• ~45 s wall-clock, fully async

In a heavy week (say 20 compacts, so ~50 summarizer calls counting L1/L2/L3/preempt): ~1-2M plan-quota tokens. Well under the limits of a Max plan.

Slash commands

You rarely need these — amnesia works on its own. They are power-user escapes.