Build Memory Systems That Make Agents Sharper Overnight

Stop re-explaining the same rules every session. Set up filesystem memory + an offline 'dreaming' pass so agents compound learning across runs.

I kept watching agents forget the same correction. One session they learn "never use em-dashes in this client’s docs." Next session they do it again because the memory lived only in that chat window. The Anthropic talk I watched this week (Lamis Mukta at AI Native DevCon) showed the actual system their Applied AI team runs: filesystem memory plus an offline curation pass they call dreaming. That’s the skill.

// The Takeaway
Treat agent memory like infrastructure, not context. Store it as plain files the agent can read and write, protect it with versioning and hash-based concurrency, then run a scheduled offline "dreaming" pass that reviews many transcripts, extracts cross-session patterns, and proposes a cleaned memory store for the next run.

Most teams still treat memory as either a long system prompt or a single MEMORY.md that grows until it is ignored. Anthropic moved through that stage: CLAUDE.md → memory tools → Skills with progressive disclosure → filesystem-as-memory (agents just use bash/grep on markdown files). Then they hit the real limit: the same agent cannot both finish the task and curate its own memory well. In-band memory splits focus and cannot see patterns across many sessions.

// The Real Shift
Separate the real-time work from the memory hygiene. The agent does the job. A dedicated overnight process reviews the transcripts and updates the shared memory.

They store team memory as markdown files in a directory. Agents read and write them with ordinary tools. Production guardrails sit on top:

Then dreaming: a batch asynchronous process that takes the current memory store plus many session transcripts (including tool calls), runs review agents to find recurring mistakes and useful workflows, and proposes an updated store with evidence. Human (or controlled) accept/reject. The next day’s agents start with higher-signal context. It does not update model weights. It is just better memory.

Slides from the talk showed teams reporting large drops in first-pass errors once this was in place.

That is the whole loop. The offline pass is what turns one-off corrections into lasting skill.

Dreaming can promote a bad pattern if you do not review the proposals. Transcripts contain tool output and user content the agent did not author, so provenance matters. This does not replace evaluation or human oversight. It is memory engineering, not magic self-improvement of the model.

Agents are staying on tasks longer and teams are running more of them in parallel. Without engineered memory, every new session starts from zero and the same mistakes repeat. The companies shipping durable agents are treating memory the same way they treat code: versioned, reviewed, and improved offline.