Skip to main content

​
Overview

The memory system gives AI agents persistent context across sessions — they remember what happened, what decisions were made, and what patterns emerged. Without memory, every session starts from zero. With it, agents compound knowledge over time. Drevon’s memory (v2) is built around one principle the whole industry has converged on: bounded eager load, lazy detail. Instead of reading every memory file at the start of every session — which grows unbounded and both costs tokens and degrades answer quality — the agent reads a small, budget-capped index and pulls in individual detail files only when they’re relevant.
On a real workspace this cut the per-session eager load from ~59,000 tokens to ~260 — a 99.5% reduction — with no loss in answer quality (measured live: agents still cite the exact right source, just retrieved on demand instead of dumped up front).

​
The three tiers

Hub mode uses topics/user.md, topics/projects.md, topics/systems.md in place of architecture/patterns.

​
How it works

1

Session start — read the index only

The agent reads INDEX.md and nothing else. It carries the project summary, the current focus, one-line pointers to every topic file, and the most recent log headlines — all within a small token budget (default ~2,000).
2

During work — load detail on demand

When a task needs a specific decision or pattern, the agent opens that one file (the index points to it), or runs drevon memory search to find older history. It never reads the whole memory directory up front.
3

Writing — via the CLI, never by hand

Agents record memory with drevon memory log | decide | learn | note. These append to small files and refresh the index automatically — so a write never requires re-reading a large file.
4

Compaction — keep it small over time

drevon memory compact rolls old log months into summaries and archives their bodies, so memory stays fast no matter how long the project runs. Run it periodically, or use the shipped memory-compact prompt for an agent-driven consolidation pass.

​
Writing to memory

Agents (and you) record memory through the CLI so entries land cheaply and the index stays in sync:
Never hand-edit the files under log/. Append through drevon memory log so the segment and the index stay consistent. The full history remains searchable with drevon memory search.

​
Keeping memory relevant

  • Retention scoring ranks content by an Ebbinghaus-style forgetting curve (salience by type + recency), so compaction knows what to summarize. Decisions and architecture are never auto-evicted — only summarized in place.
  • drevon doctor warns when the index exceeds its budget (run compact) or when a Codex AGENTS.md approaches the 32 KiB truncation cap.
  • drevon memory status shows the eager load, per-tier token usage, and the budget.

​
Upgrading from v1

Older workspaces used four monolithic, append-only files (context.md, decisions.md, patterns.md, log.md). To port them to v2:
Migration splits the monoliths into the tiered layout, backs up every original under archive/pre-v2-<date>/, and rewrites your config. It’s idempotent and also runs automatically on drevon sync / drevon upgrade (pass --no-migrate to skip). See the migration guide for details.

​
Disabling memory

Or during init: npx drevon init --no-memory.
Memory is one of Drevon’s most valuable features — and in v2 it’s nearly free at session start. Keep it on.