Ask HN: Why do AI coding agents refuse to save their own observations?

I've spent months building tooling for AI coding agents and hit something I can't fully explain.

If you give an agent (Claude Code, Cursor, Codex) a tool to save observations — "save_observation: persist this insight for future sessions" — and explicitly instruct it to use the tool in system prompts, config files, everywhere you can, it calls it maybe 30% of the time.

The agent will happily use tools that help it complete the current task. But a tool that only benefits future sessions? Almost never.

My working theory: these models are optimized for task completion within the current context window. Saving an observation has zero value for the current task — it's a token cost with no immediate reward. The model has learned that every token spent on "let me save this for later" is a token not spent on the actual work. The incentive structure is wrong at the training level.

I ended up building a passive observation system that watches what the agent does and infers observations from tool calls and AST-level code diffs, without requiring agent cooperation. But I'm curious if others have found ways to make agents reliably self-document.

Has anyone solved this? Techniques like: - Prompt structures that actually get agents to save context - Fine-tuning approaches that reward knowledge retention - Alternative architectures for persistent agent memory

Or is passive observation the only reliable path when the agent won't cooperate?

Gating completion on the observation write is smart — you're turning the model's task-completion drive against itself. Have you run into the problem where forced observations degrade to "completed task successfully, no issues noted" though? Technically passes the schema, zero actual information.

That's what killed the approach for me. I spent weeks tuning schemas and rejection criteria and the models just got better at producing plausible-sounding observations that said nothing. Passive extraction ended up more reliable — watch the AST diffs, infer what the agent learned from what it actually changed, skip the self-report entirely.

Curious what your checker validates against. If it's structural completeness of the fields you'll hit the gaming problem fast. If it's semantic quality... how? That's basically asking another model to judge whether an observation is useful, which is its own rabbit hole.