Recall: Fully‑Local Project Memory for Claude Code

What does Recall do?

Recall runs alongside Claude Code when you use the local, subscription‑based version. After every interaction it creates two files in a hidden .recall/ folder:

• history.md – an immutable, append‑only log of prompts, responses, and file edits.
• context.md – a freshly generated, human‑readable summary of the project’s current state, goals, next steps, and touched files. The summary is produced by a classical Python summarizer, so it consumes zero model tokens.

When you start a new Claude Code session, the plugin automatically injects context.md as the initial prompt. The model therefore receives a concise 1–2 K token snapshot instead of the full transcript.

Why does the context window matter?

Claude Code’s context window is limited (≈100 K tokens). Feeding the entire session history can:

1. Exhaust the window, forcing truncation.
2. Consume valuable subscription credits.
3. Require developers to rewrite project background each time.

Recall solves this by keeping only the essential information in context.md, which always fits comfortably within the window.

Business‑focused workflow

Value proposition for enterprises

• Cost control – Summarization uses no model tokens and the reduced prompt size saves credits for actual code generation.
• Speed – Developers spend seconds loading a project instead of minutes drafting a recap, shortening delivery cycles.
• Privacy compliance – All logs stay on‑premise, simplifying GDPR, HIPAA, or IP safeguards.
• Audit trail – history.md provides a complete, immutable transcript; context.md offers a concise, human‑readable snapshot for reviews.

Known limitations

• The summarizer is rule‑based Python, not an LLM; it captures high‑level changes but may miss subtle intent.
• Recall only works with locally installed Claude Code; cloud‑only or API‑driven workflows are unsupported.
• Privacy depends on the security of the host machine; teams must manage backups and access controls.

Contrarian insight

Most “memory” tools ship a cloud endpoint that ingests the full transcript, incurs per‑token fees, and expands the privacy surface. Recall demonstrates that a deterministic, locally‑run summarizer is sufficient for most development cycles. The trade‑off—potentially less nuanced summaries—is acceptable when privacy and zero‑cost operation are paramount.

Quick implementation checklist

1. Confirm Claude Code runs locally under a paid subscription.
2. Add the Recall repository to the project root (creates .recall/).
3. Include .recall/ in your internal backup routine.
4. Treat context.md as the canonical hand‑off document for new developers.
5. Periodically review history.md for edge‑case summarizer failures.

Bottom line: Recall converts the context‑window limitation from a blocker into a non‑issue for Claude Code users. By logging sessions locally and generating a compact, privacy‑first summary, it saves token credits, accelerates onboarding, and keeps confidential code on‑premise—without any additional subscription cost.

Decide what information should persist before adding project memory. The DEX capabilities overview is a useful reference for mapping that information to the actual systems an automation will use.