Self Evolve

self-evolve

An agent that doesn't learn repeats its mistakes at full price. This skill is the improvement loop: observe the run → distill the lesson → store it → reinforce what worked → recall it next time. It uses local tools (engram for memory, tracelet for observation), so the learning stays on the user's machine.

The loop (run it at the end of each non-trivial task)

1. Observe what actually happened. Don't trust your memory of the run — check it. If tracing is on (npx @jnmetacode/tracelet), look at the trace: which tool calls failed or were retried, where the time and tokens went, which approach was abandoned. Without a trace, review your own steps: what did you try first that didn't work?

2. Distill at most 1–3 lessons. A lesson is durable knowledge that would change how you act next time, not a diary entry:

   • a root cause ("the staging DB rejects connections without SSL")
   • a working recipe ("build fails unless npm ci runs before codegen")
   • a user preference revealed by a correction ("they want diffs, not files") Skip anything the repo/docs already record.

3. Store each lesson (one fact per memory, dated, with the why):

   engram_remember: "2026-06-12: deploys to staging need SSL_MODE=require —
   the pooler silently drops non-SSL connections (cost us 40 min)."
   

   (CLI: npx @jnmetacode/engram remember … via the HTTP API, or the engram_remember MCP tool.)

4. Reinforce retrievals that proved right. If you recalled a memory during the task and it turned out to be the correct answer, say so — recall gets measurably better with use:

   engram_reinforce: { query: "staging deploy fails", source: "deploys" }
   

   Future similar queries will rank that source higher. Only reinforce what was verified (the fix worked, the user confirmed) — reinforcing guesses trains the memory to be confidently wrong.

5. Close the loop next task: recall before you act. Start non-trivial tasks with a recall of the relevant area (engram_recall: "staging deploy"). The loop only pays off if stored lessons actually get read.

Calibration

• Don't hoard. One sharp lesson beats ten vague ones; if everything is a lesson, nothing is. When a memory turns out to be wrong, delete it (engram forget <source>) rather than stacking corrections on top.
• Promote recurring lessons into skills. The third time a lesson is about procedure ("always do X before Y in this kind of task"), it has outgrown memory — write it into a SKILL.md (npx @jnmetacode/skillet new <name>) so every future session starts with it instead of having to recall it.
• Stay honest. The loop measures itself: if reinforced recalls keep surfacing the wrong source, the lessons are badly written — rewrite them more concretely instead of reinforcing harder.