Cross-Session Learning: Accumulation Effects

+15 Mana ✨
Hermes long-running session behavior(DOCUMENTATION)

Introduction

The value of user profiling is not what one session adds; it is what twenty sessions accumulate. Each session writes maybe one or two new entries, refines an old one, or removes a stale one. After a few weeks, USER.md reflects how you actually work, not how you described yourself once. Understanding this slow accumulation shapes how you use the agent.

Key Concepts

  • Per-session delta: Profiling typically produces a small change per session, not a flood.
  • Refinement: Entries get rewritten as the agent learns more precise language.
  • Stability over time: After a few weeks, the rate of change slows; the profile converges.
  • Curation moments: Periodic human review accelerates and corrects accumulation.

Real World Context

A developer enables profiling on a fresh install. After session 1, USER.md has one entry inferred from setup banter. By session 10, it has four entries, two of which the agent has consolidated. By session 30, the file is roughly steady-state with five to seven entries that genuinely describe how the developer works. The first session felt magical. The accumulated thirtieth session feels normal because it just works.

Deep Dive

A few accumulation dynamics to know:

  1. The first few sessions are noisy: The agent has little to anchor on, so early entries are sometimes too broad. This is normal.
  2. Repetition causes refinement: Telling the agent "please be terse" three times across sessions causes it to either consolidate existing entries or replace a weaker entry with a stronger one.
  3. Contradictions cause replacement: If a profile entry says "prefers verbose explanations" and the user starts demanding terse answers, the agent will replace the old entry rather than keeping both.
  4. Stability emerges: Once the high-leverage facts are written, the agent slows down on new writes. A steady-state USER.md is a sign of a working profile.

A common surprise: the accumulated profile is sometimes better than what you would have written yourself, because it captures behaviors you do not consciously notice. You may not realize you always run tests after edits, but the agent will write "runs tests after every edit" once it sees the pattern five times.

The inverse is also true: the agent can encode a pattern that was a phase, not an identity. If you spent two weeks in a Rust project and the profile now says "prefers Rust", you may want to correct it back to "comfortable in Rust, primary work in Go and TypeScript." Human curation is how you keep the profile honest as you change.

Common Pitfalls

  1. Expecting the profile to be useful immediately: Profiling rewards patience. The first session is rough; the tenth is good; the thirtieth is great.
  2. Letting transient phases harden into permanent entries: A two-week project should not rewrite the identity layer. Curate when phases end.

Best Practices

  1. Resist the urge to over-correct early: A weird entry in week one is not a problem. Refine across sessions or wait for the agent to consolidate.
  2. Schedule a quarterly read: Open USER.md every few months. Two minutes of editing keeps the profile aligned with current reality.

Summary

  • User profiling compounds across sessions; per-session deltas are small.
  • Early sessions are noisy; the profile stabilizes after a few weeks of use.
  • Repetition and contradictions drive refinement and replacement.
  • Human curation corrects for phases that should not be encoded as permanent identity.

Code Examples

text
Session 1  →  USER.md: 1 entry  ("User is a developer")
Session 5  →  USER.md: 3 entries (role, comm style, timezone)
Session 15 →  USER.md: 6 entries, 1 consolidated
Session 30 →  USER.md: 5-7 entries, mostly stable
Session 60 →  USER.md: 5-7 entries, near steady-state

# The shape is logarithmic: fast growth at first, then convergence.
✓ Completed