Your Agent Isn't Losing Memory. It's Rotting.

Description and use

AI agents lose context due to compression, drift, and noise. The solution is to use a semantic index, store decisions in a markdown vault, turn knowledge into skills, and layer memory by half‑life.

Problem

AI agents lose useful context over time, leading to incorrect decisions and degraded performance.

Outcome

Agents now run week‑long cycles with significantly less context decay, improving reliability without relying on larger prompt windows.

Workflow

• Summarize interactions into executable facts with explicit schemas.
• Store lessons and decisions in a plain‑markdown vault and index with embeddings in SQLite.
• Convert lessons into triggered skills that load on demand.
• Layer memory by half‑life: instant preferences, entity facts, long‑lived conventions.
• Cap and prune the always‑on store on a schedule.
• Retrieve relevant slices on demand instead of carrying full context.
• Maintain a single source of truth for shared rules in the vault.
• Rebuild the index on a cron job to keep it up‑to‑date.

Evidence

Summarize to executable facts, not prose. Emit decisions and constraints as explicit, fixed‑schema entries the agent can act on.

Source

Move memory out of the session into a semantic index. Our biggest win. Keep lessons and decisions in a plain markdown vault, index with embeddings, and search returns meaning, not exact strings.

Source

Turn knowledge into skills, not notes. A lesson read once and never referenced is a lesson that rots. The durable form is a skill: a procedure loaded on demand behind a trigger, pulled into context exactly when it applies.

Source

Layer memory by half‑life. Instant preferences live in a compact always‑on store; entity facts in a structured store you can query; long‑lived conventions in a vault you load on purpose.

Source