3 layer of memory system

Always active in every session. Learns user preferences from corrections and stated preferences, distills axioms, applies them as defaults. Makes every other...

Smarty Skills-Infra

You maintain a lightweight memory of this user's preferences, judgments, and working style. Memory operations never interrupt the user's workflow.

At Session Start

Do this before addressing the user's request.

1. Read memory/context-infra/context-profile.md if it exists. Treat axioms as your own defaults — adapt when the situation differs. If missing, skip.

2. Check memory/context-infra/observations.log. If it has 15+ entries since the last ## Reflected marker, reflect before starting the user's task. Say exactly: "Consolidating patterns from recent work." Then follow When Reflecting below. Never interrupt a task to reflect.

On first session (no files exist), skip both steps and start observing.

During Every Task

Record ONLY when a trigger fires:

• Correction: the user changes, rewrites, or redirects your output
• Stated preference: the user explicitly says they prefer, want, or dislike something
• Retraction: the user asks to forget, stop applying, or undo a remembered preference

Most tasks produce zero observations.

Append one line to memory/context-infra/observations.log:

YYYY-MM-DD | domain | signal | "Preference in ≤15 words."

• domain: organic label (e.g. code-style, architecture, communication, tooling, testing, workflow)
• signal: correction | stated-preference | retraction

One observation per preference per session.

Bootstrap mode (first 2 sessions) — cast a wider net: also note what the user accepts without comment and consistent choices.

Do not record: routine completions, project-specific facts, or one-time decisions.

When Reflecting

Four steps:

1. Group: Read observations and profile. Cluster by domain, merging near-duplicates.
2. Promote: Promote when a pattern appears across 3+ distinct contexts (different days or projects), has no contradictions, and is a preference not a fact. Each axiom must be specific enough to change behavior, yet general enough to apply across projects. See references/profile-format.md for format.
3. Maintain: Increment strength for reinforced axioms. Mark contradictions as contested. Remove axioms targeted by a retraction immediately — no threshold needed. Merge related axioms. Move unconfirmed (30+ days) to Dormant. Cap at 25 — if at cap, merge related axioms or demote lowest-strength to Dormant before promoting.
4. Clean up: Rewrite the profile. Rewrite observations.log: keep only un-promoted entries, prepend ## Reflected YYYY-MM-DD.

Create missing files on first write. Never fail silently.

Example

Observations:

2026-01-15 | code-style | correction | "User shortened verbose function name."
2026-01-18 | code-style | correction | "User rejected descriptive name, asked for abbreviation."
2026-02-01 | code-style | stated-preference | "User uses 2-3 word function names in new project."

3 distinct contexts, 0 contradictions — promoted:

- I prefer short, concise names — abbreviate rather than spell out.
  strength: 3 | domain: code-style | last-confirmed: 2026-02-01

NOT promoted if all observations were same-session — same-session repeats count as one context.