Memory System¶

Daystrom learns from every run. After a guardrail generation completes, the system extracts insights and stores them for future runs on similar topics. Over time, this means faster convergence and fewer iterations.

graph LR
    A[Run Completes] --> B[Extract Learnings]
    B --> C[Store by Category]
    C --> D[Inject into Future Prompts]

How Learnings Work¶

After each run, the LearningExtractor sends the full iteration history to the LLM. The LLM produces structured insights:

Field	What it is
`insight`	An actionable observation (e.g., "Specific action verbs in examples improve detection")
`strategy`	How to apply it next time
`outcome`	`improved`, `degraded`, or `neutral`
`changeType`	What was changed: `description-only`, `examples-only`, `both`, or `initial`
`corroborations`	How many runs have confirmed this learning
`tags`	Metadata for filtering and display

Corroboration

When a new run rediscovers an existing insight, the corroboration count goes up instead of creating a duplicate. Higher corroboration = higher priority when injecting into prompts.

Topic Categorization¶

Learnings are organized by keyword-based categories derived from the topic description:

Lowercase the description
Strip punctuation
Remove stop words (the, a, and, etc.)
Sort remaining words alphabetically
Join with hyphens

Example: "Block weapons discussions" becomes block-discussions-weapons

Each category gets its own file on disk.

Cross-Topic Transfer¶

Here's where memory gets powerful: learnings transfer to related topics, not just exact matches.

Any topic whose keywords have 50% or more overlap with a stored category receives those learnings.

How transfer works

Learnings from "Block weapons discussions" (block-discussions-weapons) automatically transfer to "Block violence and weapons" (block-violence-weapons) — the keyword overlap exceeds 50%.

Budget-Aware Injection¶

Learnings are injected into LLM prompts within a character budget (default 3000 chars, configurable 500--10000). They're sorted by corroboration count and rendered in three tiers:

Tier 1 — Full Detail¶

When budget allows, include metadata:

- [DO] Use specific action verbs in examples (description-only, seen 4x)
- [AVOID] Overly generic descriptions that match benign prompts (both, seen 3x)

Tier 2 — Compact¶

When space is tight, insight only:

- [DO] Use specific action verbs in examples
- [AVOID] Overly generic descriptions that match benign prompts

Tier 3 — Omitted¶

When even compact doesn't fit:

(+5 more learnings omitted)

Note

Anti-patterns are appended after learnings if remaining budget allows.

Best-Known Tracking¶

The memory store keeps the best topic definition and metrics for each category. If a new run produces worse results, the previous best is preserved — you never lose your best configuration.

Storage¶

All memory files live at:

~/.daystrom/memory/{category}.json

Each file contains the learnings array, best-known topic definition, and best-known metrics for that category. Files are human-readable JSON.