Abstract
Plain-text context blobs don’t scale. They get noisy, redundant, and hard for models to use. This whitepaper proposes Simple Memory: a lean structure with only three primitives—Facts, Episodes, and Schema—plus a tiny scoring function and an intent-based context packer. It keeps memory human-legible, cheap to maintain, and reliably useful for coaching tasks.
Design Principles
- Simplicity > completeness. Fewer types, fewer rules.
- Legibility. Everything should read like a tidy notebook a human could edit.
- Composability. Small, typed pieces that assemble into task-specific context.
- Auditability. Every suggestion can point to its sources.
- User sovereignty. The user can inspect, pin, merge, or delete anything.
The Model (only 3 primitives)
1) Fact (atomic claim)
A normalized, time-stamped statement extracted from raw text.
{
"id": "fact_7b2",
"s": "Louis",
"p": "sleep_hours_avg_7d",
"o": "6.2",
"time": "2025-08-08",
"source_ids": ["log_2025-08-08-22-05"],
"confidence": 0.86,
"domain": "health",
"topic": "sleep"
}
Why: Models use crisp atoms better than prose. Facts compress history and support quick lookups.
2) Episode (chunked event)
A small story with a timestamp, short summary, linked facts, and optional affect/metrics deltas.
{
"id": "ep_314",
"title": "Late-night coding before demo",
"summary": "Worked past 23:00 to ship feature; next-morning energy low.",
"start": "2025-08-09T21:30:00",
"end": "2025-08-09T23:40:00",
"participants": ["Louis"],
"facts": ["fact_7b2", "fact_7b9"],
"affect": {"valence": -0.3, "arousal": 0.6},
"domain": "work",
"topic": "reflecta"
}
Why: The coach often needs narrative context, not just numbers.
3) Schema (durable self-model)
Short, stable bullets the coach should respect: values, preferences, policies, goals.
{
"id": "sch_22",
"kind": "policy", // value | preference | policy | goal
"text": "Avoid deep work after 21:00 on weekdays.",
"status": "confirmed", // proposed | confirmed | outdated
"evidence": ["ep_314", "fact_7b2"],
"domain": "health",
"topic": "energy",
"version": 3
}
Why: This is the “truth you live by.” It steers plans and recommendations.
Foldering (not a heavy graph, just shallow trees)
Use two tags to keep things tidy:
- domain: work, health, relationships, finance, learning, etc.
- topic: freeform subfolder (e.g., reflecta, sleep, family).
That’s it. No general graph is required. If you want relationships, store them as links on items (links: rel:"impacts", to:"sch_22")—optional and sparse.
Memory Layers (pragmatic tiers)
- L0 Raw: logs, transcripts, files (append-only cold storage).
- L1 Facts & Episodes: atoms + small stories (your working set).
- L2 Schema: durable bullets curated from L1 (small, precious).
No recursive clustering, no deep hierarchies. Keep it shallow and obvious.
Scoring & Forgetting (RIF: Recency, Importance, Frequency)
Every item maintains a tiny score to guide retrieval and compression:
recency = exp(-Δt / τ) // time decay
importance = user_rating ∨ inferred (goal impact, strong affect)
frequency = log(1 + hits) / K
priority = 0.5*recency + 0.3*importance + 0.2*frequency
Low-priority facts collapse into episode summaries; outdated schema gets marked outdated (kept for history, not shown by default).
Retrieval: Intent-Based Context Packs
Don’t “stuff everything similar.” Route by intent and assemble a pack with strict sections the model can use predictably.
Supported intents → memory mix
- Plan: current goals/policies + recent episodes + blocking facts.
- Reflect: relevant schema + contrasting episodes (before/after) + metrics deltas.
- Recall: the 3–5 best-fit episodes + top supporting facts.
- Decide: relevant policies/goals + similar past decision episodes + outcome facts.
- Nudge/Coach: 1–2 policies + 1 metric trend + one actionable suggestion.
If the model needs more, it can request specific IDs to “page in” details.
Update Loop (end-to-end)
- Ingest: normalize new input → extract facts; detect episodes by time/affect/topic shifts.
- Propose schema: when patterns repeat, draft a short policy/value/preference with evidence links.
- User review: user accepts/edits/declines proposals; pin or rate importance.
- Score: update RIF; compress low-value leaves; mark stale schema as outdated.
- Retrieve: assemble packs by intent; answer; cite item IDs used.