Consolidation and forgetting in memory systems

Episodic memory grows without limit. Without consolidation it becomes a heap of events nobody can use.

Consolidation is the job that periodically reads the new episodes and updates a short list of facts: «prefers code examples», «is working towards the LoRA project by December», «knows linear algebra».

Three rules that decide the quality:

1. Support: a fact requires at least k episodes (or an explicit statement from the user). Otherwise isolated coincidences become «truths».
2. Confidence: every fact carries a number that is raised on new support and falls with time.
3. Validity: facts that are time-bound («revising for the May exam») get an expiry date.

Forgetting is not a shortcoming — it is what keeps the list usable.

import time, math

HALF_LIFE_DAYS = {"preference": 180, "knowledge": 365, "goal": 90, "temporary": 14}
MIN_CONFIDENCE, MIN_SUPPORT = 0.3, 2

def consolidate(facts: list[dict], new_episodes: list[dict], llm) -> list[dict]:
    candidates = llm.extract_facts(new_episodes)          # [{text, type, sources}]
    index = {f["text"].lower(): f for f in facts}
    now = time.time()
    for c in candidates:
        key = c["text"].lower()
        if key in index:                                   # reinforce an existing one
            f = index[key]
            f["sources"] = sorted(set(f["sources"]) | set(c["sources"]))
            f["confidence"] = min(1.0, f["confidence"] + 0.25)
            f["last_supported"] = now
        elif len(c["sources"]) >= MIN_SUPPORT:             # a new fact requires support
            index[key] = {**c, "confidence": 0.5, "last_supported": now}
    return forget(list(index.values()), now)

def forget(facts, now):
    kept = []
    for f in facts:
        age_d = (now - f["last_supported"]) / 86400
        hl = HALF_LIFE_DAYS.get(f.get("type", "preference"), 180)
        f["confidence"] *= 0.5 ** (age_d / hl)
        if f["confidence"] >= MIN_CONFIDENCE:
            kept.append(f)
    return sorted(kept, key=lambda f: -f["confidence"])[:50]      # a hard ceiling

Measure the precision of the list of facts: take a sample of 30 facts and let a human being (or the user themselves) mark them correct / incorrect / out of date. Below 90 % precision the memory does more harm than good — then the thresholds should be raised.