Memory

A working memory system

A practical guide to keeping Claude useful across sessions — covering the session memory files, project context patterns, and retrieval habits that prevent Claude from starting cold every time you open a new chat.

If you have noticed that every session loses what the last one built, this gives you a concrete system for fixing that.

Five areas, one persistent system

• session-memory — Memory files Claude loads at session start: anchor-formatted Markdown with a priority header, an active-projects list, and a standing-rules section. Token budget: keep the primary memory file under 4k tokens so it always fits in the pre-prompt without displacing working context.
• project-context — Keeping Claude oriented on long-running work: a CLAUDE.md PROJECT: section updated after every significant session, a decisions log with dated entries, and a session-loader pattern that reads context in order (global rules → project state → today’s brief).
• semantic-retrieval — Surfacing past context at the right moment. Two patterns: embedding-based similarity search over memory files (vector store, top-k retrieval) and keyword fallback over a structured JSONL log. The hybrid approach that uses keywords for speed and embeddings for recall.
• cross-session-continuity — Carrying decisions and state across days of real work: a decision log with four fields (date, decision, rationale, supersedes), a versioned rules file (each rule has an ID and a date), and a handoff brief that summarizes open threads at the end of every session.
• context-prioritization — What belongs in memory vs. what to reconstruct on demand. Decision heuristic: if reconstructing it costs more than loading it, load it. Token cost calculation, stale-detection by last-modified date, and a decay policy (hard facts stay; task status expires in 7 days).
• project-context — Keeping Claude oriented on long-running projects without pasting the whole history every time.
• semantic-retrieval — How to surface the right past context at the right moment rather than searching manually.
• cross-session-continuity — Patterns for carrying decisions, rules, and state across days of real work.
• context-prioritization — How to decide what belongs in memory and what’s cheaper to reconstruct on demand.

A preview

The problem isn’t that Claude forgets. The problem is that every session starts as if the last one never happened. Memory files fix that — not by making Claude smarter, but by making sure it starts informed.

A session memory file isn’t documentation. It’s a briefing — the minimum Claude needs to pick up where you left off without you re-explaining everything.

Intermediate users

Intermediate Claude users who have noticed that every session starts cold and want a working system to fix that. No prerequisites — works alongside any existing Claude workflow.

A longer excerpt is available to newsletter subscribers.

Buy on Amazon