🧠 Recall
Get the most out of Claude Code's 200K context window — without paying for the 1M token tier.
Claude Code's 200K context sounds like a lot, but ~20-25% goes to system overhead (tools, prompts, CLAUDE.md), and auto-compaction triggers at ~85%. That leaves roughly 120-130K usable tokens — and the 1M tier isn't 5x more, it's effectively ~8x the usable context (since overhead is fixed).
Recall bridges this gap by making the 200K window renewable: when context runs out, start a new session and restore everything that matters in seconds.
Without Recall, compaction compresses your entire conversation into a summary. Recall preserves far more detail (15-18K tokens of actual conversation), giving Claude the full conversation arc to continue where you left off.
Why Recall Exists
Claude Code's built-in compaction has gotten better over time — it captures the main topics, key decisions, and important files. But a summary is inherently lossy. The longer and more nuanced the session, the more detail gets compressed away:
| What a summary loses | Why it matters |
|---|
| Exact phrasing of user instructions | Claude may interpret your intent differently than you stated it |
| Failed approaches & why they were abandoned | Claude might revisit something you already ruled out |
| Architecture decisions & rationale | The reasoning behind a choice gets flattened into just the outcome |
| Intermediate debugging steps | A related issue later may need that exact context |
| File edit history & frequency | Claude loses track of what it already changed and how often |
| Code review feedback & corrections | Your specific preferences don't carry over |
Recall keeps the actual conversation instead of a summary — every user message, every assistant response, every tool call. A summary can never fully replace the real thing, and when you're deep in complex work, the details matter.
What Recall preserves
Recall parses the complete session transcript — including context from before any previous compactions in the same session. Even after multiple compactions, nothing is permanently lost. 99%+ of noise is stripped (tool results, thinking blocks, system reminders) while keeping:
- Every user message — verbatim
- Every assistant response — verbatim
- Tool call summaries — which tools were called with what parameters
- Compaction markers — so you can see where context was previously lost
- Loaded skills — so they can be re-loaded after compaction
See the difference
Both examples below show the same fictional session — adding recurring tasks to a Swift app:
-
Compaction summary (~3,500 tokens) — What Claude Code produces by default. Preserves the topic and key files, but not the details: that you debated UTC vs. local-time storage and chose UTC, that weekday/weekend presets were added, or that a scheduling bug was caused by computing from Date() instead of scheduledDate.
-
Recall transcript (~15,000 tokens) — What Recall produces. Every user message verbatim, every assistant response, every tool call summarized. The full conversation arc with decisions, bugs, fixes, and rationale.
Installation
Step 1: Install the plugin
Start Claude Code, then run:
/plugin marketplace add FlineDev/Marketplace
/plugin install recall
If you're in an active session, run /reload-plugins to activate immediately. Recall is part of the FlineDev Marketplace — see the full list of available plugins there.
[!TIP]
Automatic Updates: By default, third-party plugins don't auto-update. To receive new features and fixes:
- Type
/plugin and press Enter
- Switch to the Marketplaces tab
- Navigate to FlineDev and press Enter
- Press Enter on Enable auto-update
Step 2: Start using Recall
After installation, /recall:session works immediately — no per-project setup needed:
/recall:session <session-id>
Start a new session, paste the session ID from a previous chat, and Recall restores the full conversation context. That's it.
Optional: Automatic compaction integration
If you want Recall to automatically kick in whenever compaction happens (auto or manual /compact), run this once per project:
/recall:compact-on
This hooks Recall into the compaction lifecycle so recovery happens without you having to do anything. To remove the integration later, run /recall:compact-off.
Requirements
- Claude Code (v2.0.76+ recommended)
- Python 3 (pre-installed on macOS)
Usage
Recall covers both ways people use Claude Code:
With auto-compaction (default)
This is the "set it and forget it" mode. You don't change anything about how you work. When auto-compaction triggers — or when you manually run /compact — Recall kicks in automatically.