Memory Is Files

AI has a memory problem. Every conversation starts from zero. Context windows fill up and get compacted, losing detail. Switch to a different model or a different tool and your history doesn't follow. The industry is building elaborate solutions — vector databases, knowledge graphs, retrieval-augmented generation, memory layers that compress and index and embed. These are interesting engineering. They are also, for most practical purposes, overkill. The simplest solution to AI's memory problem is a file.

A worklog is a plain text file — markdown, usually — that records what was done, what was found, and what was concluded. Not a transcript of the conversation. Not a compacted summary. A running record of action and result, written by the AI as it works. When you're investigating something across multiple sessions — digging through logs, testing hypotheses, building evidence toward a conclusion — the worklog is the thing that survives between sessions. You open a new conversation, point the AI at the worklog, and you're back to full speed. No re-explaining. No losing the thread. No hoping the compacted summary preserved the thing that mattered.

The key: the AI writes the worklog, not you. This keeps the AI honest. Every claim, every finding, every decision is documented in a file you can read, verify, and grep. The AI can't quietly lose context or silently revise its understanding, because the log is right there. And it's full of context — the AI was present for every detail of the session, so nothing gets lost to the human tendency to summarize too aggressively or forget the thing that seemed minor at the time but turned out to matter.

This works because of what a worklog filters. A conversation with AI is full of dead ends, misunderstandings, corrections, tangents, and restatements. The worklog strips all of that away and keeps the signal: this is what we tried, this is what we found, this is what it means. The AI does the compressing, but the human directs the work — which means the human decides what matters by deciding what to investigate. The AI preserves the full detail of the investigation. It's a partnership: human judgment on what to pursue, AI diligence on what happened.

The worklog also solves the multi-agent problem. If you're using one tool for code, another for research, and a third for writing, none of them share context with each other. But they can all read a file. Update the worklog in one session, open it in the next, and the new agent inherits the full state of the project regardless of which model or tool is running it. The file becomes the shared memory layer — not because it's architecturally sophisticated, but because every AI tool that exists can read a markdown file.

Over time, the worklog becomes more than a memory aid. It becomes a searchable record of how a project unfolded. You can grep it. You can skim it for patterns. You can hand it to someone else — human or AI — and they can understand not just where the project stands but how it got there. This is something conversation histories and vector embeddings can't do, because they preserve the raw material without the editorial judgment. The worklog is already edited. It's already a narrative of what mattered.

There is a specific discipline to writing a useful worklog. Date your entries. Record what you did, not what you planned to do. Note dead ends explicitly — they're valuable because they prevent the next session from retreading the same ground. Keep it concise but specific: "Tested hypothesis X against dataset Y, result was Z, which means W" is a useful entry. "Worked on the project" is not. And update it in real time, during the work, not after — because the thing you'll forget to write down at the end of the session is exactly the thing you'll need to remember next time.

Scale it up and you get dossiers. A worklog tracks a project. But if you tag the facts in those worklogs to entities — people, places, projects, deadlines — and connect those entities with relationships, the accumulated observations form a graph. Over time, the graph fills in. The AI mentions Aaron in five different worklogs across three weeks, and now the system knows Aaron's role, his projects, his schedule, his preferences — not because anyone sat down and wrote a profile, but because the observations accumulated on a skeleton of relationships. The dossier grows from conversation. The graph gives it structure.

The people building memory systems for AI are trying to solve this problem with infrastructure. Embedding databases, context repositories, hierarchical summarization. Some of this work is genuinely important and will matter at scale. But for an individual working with AI on real projects across days and weeks, the answer is already here. It's a file. It's been a file this whole time. Memory is files.