Agents ignored the prompt to read memory in 7 of 7 sessions: an analysis of 13 thousand logs

The kaeru team, which builds long-term memory for AI agents such as Claude Code, Codex and Cursor, has published an analysis of its own logs: more than 13 thousand working-session files from June through September. The headline result: in the 7 sessions where an agent saw the prompt "you can read the saved reasoning," it never read it, even though it followed another prompt, "link this record to the previous one," 751 times. For several months the authors designed the product by feel, then checked how agents actually use memory. A lot turned out to be wrong, and most of the findings apply beyond kaeru, to anyone writing an MCP server, hooks or long instructions for agents.

What kaeru is and what was tested

kaeru is a local server that an agent connects to over MCP, the standard protocol for calling external tools. As it works, the agent writes its own records: observations, decisions, tasks with deadlines, hypotheses with test results. Records are connected by weighted links, so a whole chain of reasoning can be saved. A corrected fact is not overwritten; the old version is closed with a date. When the agent enters a project, it receives a summary: what is current, what is overdue, and what reasoning already exists. All of this works only if the agent reads memory before asking a question, records what it has found out, and marks what is outdated.

The authors are upfront about the limitations: these are their own logs, not a market-wide sample. The samples differ between sections, from thousands of events to seven sessions or the logs of a single user. They measured agent behavior, not benefit. The team does not yet know how much time or how many tokens memory saves per task, and is preparing its own benchmark.

A debt gets paid, an option gets skipped

On project entry, the team added the prompt "this project has saved reasoning, you can read it." Over four days the model saw it in 7 sessions and never read it, although in the same period agents saved 17 new pieces of reasoning. They are willing to write, but not to reread.

Another prompt, "link this record to the one it grew out of," arrived right after a record was written and was followed 751 times. It is the most frequent action in the whole product. In version 0.8 the link is specified directly in the write call, so a separate prompt is no longer needed. The difference is timing: the second prompt continues what the agent is already doing, while the first asks it to drop the task and do something unrelated.

The summary shows the same pattern. On the line "2 records in question, awaiting decision," agents worked through disputed records 8 times, and overdue tasks get closed reliably, while the line "you can read the saved reasoning" never triggered once. Everything in the summary is now phrased as a debt.

Hooks, truncated instructions and optional parameters

In one user's logs, 80% of the agent's questions to the human came with no memory access in the previous ten minutes; the median was 74 minutes since the last read. For this the team built a hook, a script the client runs at a set moment. When the agent is about to hand the turn to the human, the hook searches memory for material on the topic and sends the agent to read it once.

The first version detected a request to the human by the "?" character. Across 2 608 agent stops, it caught 201 of 1 164 requests, or 17%.

In 856 cases the requests were imperatives: "tell me," "send the report," "I need your answer on the supplier." Agents ask for things more often than they ask questions.

The most unpleasant finding: the client truncated the MCP server's instruction at 2 048 characters, and nobody noticed for two months. Among tools mentioned before the cutoff, 9% went unused. Among tools mentioned after the cutoff, 31% did, the same as for tools the instruction did not mention at all. For the model, the tail of the instruction simply did not exist. kaeru's introductory text is now 1 499 characters.

Optional parameters behaved the same way. Link weight was optional, with a default value. Across 6 003 memory tool calls, agents created 1 262 links and never once specified a weight. In 0.8 it became mandatory. For reminders, the display deadline has been mandatory from the start: "the certificate expires" and "re-measure the figure before citing it" need different deadlines, and only the author of the record knows them.

The approach that worked best was background cleanup, which asks nothing of the agent: on its first pass it sorted 33 accumulated records into layers on its own.

The team's practical takeaways come down to five rules:

• what is needed should arrive on its own, inside a response the agent receives anyway;
• phrase it as a debt: "awaiting decision," not "you can take a look";
• prompt at the moment of action, not everything at once at the start of a session;
• make required parameters without defaults;
• keep the instruction short and check how much of it actually reaches the model in your client.

kaeru's source code is published on GitHub, and the next article in the series will show what a working day looks like with these findings applied.