Phase 10 · Independent Memory Layer: episodic + semantic + traceability

Maps to report P0 "memory independent layer". Source: examples/common/memory.py, examples/p10/hello_chain.py.

Why this layer

Earlier "memory" was only a checkpoint state snapshot (MemorySaver) — a task-level snapshot, not real, durable / traceable knowledge. A production Agent needs:

• Episodic (short-term): what happened in a session, replayable;
• Semantic (long-term): facts distilled from dialogue, reusable across sessions;
• Traceability: every semantic memory carries a source; answers must cite it — the anti-hallucination switch.

Three layers

from examples.common.memory import MemoryStore
ms = MemoryStore()                        # :memory: demo; pass a file path for persistence
eid = ms.add_episode("s1", "user", "We chose LangGraph.")
ms.add_semantic("Orchestration framework = LangGraph", f"episode:{eid}", "s1")

Retrieve with citations:

ctx = ms.build_citation_context("orchestration")   # returns facts with [source]
# empty when no source -> caller must NOT let the model assert; refuse instead

Anti-hallucination

build_citation_context returns "" when nothing matches. Inject it into the prompt: non-empty → model may only answer from memory and cite; empty → model says "no memory found", no fabrication.

Run

make run p=10

Summary

• Memory ≠ checkpoint snapshot: checkpoint is task-level state; the memory layer is durable, traceable knowledge.
• The three layers (episodic / semantic / traceability) let the Agent cite sources — the hard switch for anti-hallucination.
• With memory as an independent layer, RAG handles retrieval and memory handles "known facts" — cleaner separation of concerns.