Semantic memory and consolidation
Be able to distil episodes into facts, handle contradictions and forgetting, and connect the memory to the agent's decisions.
Intuition
Episodes are events. Semantic memory is what the agent knows about the user after seeing many events: «prefers code examples to formulas», «knows linear algebra», «goal: LoRA fine-tuning by December». It is built through consolidation: periodically the model reads the new episodes and updates a list of facts.
Every fact carries: the source (which episodes), the confidence, the time, the validity (permanent / for now / expires). That makes three things possible:
- Contradictions: a new fact «wants formulas» against an old «prefers code» → the new one wins if it is fresher and supported by several episodes; otherwise ask the user.
- Forgetting: facts without support for a long time fall in confidence and are removed — otherwise the list grows and becomes wrong.
- Decisions: the agent reads the facts before it chooses the depth of explanation, the difficulty, the tone. If the facts do not affect any decision the memory is decoration.
Measure: the decisions with vs without the facts; the share of facts that are correct on human inspection; the share of contradictions that were resolved correctly.
Code
import json, time
CONSOLIDATE = """Here are known FACTS about the user (a JSON list) and NEW EPISODES.
Update the list of facts: add, raise/lower the confidence, mark contradictions. Every fact:
{"text": ..., "confidence": 0-1, "sources": [episode-id], "valid_until": null|date, "type": "preference|knowledge|goal"}
Answer ONLY with the new JSON list."""
def consolidate(llm, facts, episodes):
out = llm(CONSOLIDATE + "\nFACTS: " + json.dumps(facts, ensure_ascii=False) + "\nNEW EPISODES: " + json.dumps(episodes, ensure_ascii=False), temperature=0, json_mode=True)
new = json.loads(out)
return forget(new)
def forget(facts, half_life_d=90, min_conf=0.3):
now = time.time(); out = []
for f in facts:
age = (now - f.get("last_supported", now)) / 86400
f["confidence"] *= 0.5 ** (age / half_life_d)
if f["confidence"] >= min_conf and not (f.get("valid_until") and f["valid_until"] < time.strftime("%Y-%m-%d")):
out.append(f)
return out
def to_decision(facts):
pref = [f["text"] for f in facts if f["type"] == "preference" and f["confidence"] > 0.6]
return {"explanation_depth": "code" if any("code" in p for p in pref) else "intuition"}
Mastery means
- Distils episodes into facts with a source and a validity
- Handles contradictions and forgetting
- Connects the memory to the agent's decisions and measures the effect
Prerequisites
- EVector databases and indexingrequired
- FEpisodic memory for agentsrequired
Sign in to do the exercises and build your mastery up.
Sources
- arXiv — MemGPT: Towards LLMs as Operating Systems — arXiv (open access; licence per article)
- arXiv — Generative Agents: Interactive Simulacra of Human Behavior — arXiv (open access; licence per article)