Episodic Trace Memory for LLM Agents

The Problem

Standard RAG gives LLM agents access to facts — but not to how a problem was solved before. This maps to cognitive science’s separation between semantic memory (what you know) and episodic memory (what you’ve done and how it went).

Current vector stores are unordered — they cannot return the sequence of states an episode traversed. They have no temporal structure, no concept of episode identity. CVX has all three: entity-level trajectories, temporal ordering, and causal search.

Core Idea: Reward-Filtered Temporal Analogy

“Find past reasoning episodes that began in a state similar to my current state, and return the path they took — ranked by whether they succeeded.”

Capability	Standard Vector Store	CVX
Find similar states	Cosine similarity	Cosine + temporal distance (α-weighted)
Return ordered trajectory	Not possible	trajectory(entity_id) — ordered by timestamp
Filter by outcome	Post-hoc filtering	MetadataFilter::gte("reward", 0.5)
Show what happened next	Not possible	causal_search() — successor edges
Episode identity	Not native	`entity_id = episode_id << 16

Data Model

Each reasoning step is a TemporalPoint:

vector:     embed(state ⊕ action)     — what was the situation + what was done
entity_id:  encode(episode_id, step)  — episode:step encoded as u64
timestamp:  monotonic within episode  — step ordering
metadata:   {episode_id, step_index, reward, action_type, outcome_text}

Retrieval Pipeline

1. Phase 1 — Analogous episode discovery: snapshot_knn with metadata_filter(step_index==0, reward>=0.5) to find initial states of successful episodes
2. Phase 2 — Full trajectory retrieval: trajectory(entity_id) for each candidate episode
3. Phase 3 — Context injection: format retrieved traces into structured LLM prompt with success/failure annotations

CVX API Surface

CVX Feature	Used for
snapshot_knn + MetadataFilter	Finding initial steps of successful episodes
trajectory()	Reconstructing full episode
causal_search()	”What happened next to similar states?”
velocity()	Filtering out high-variance noisy episodes
detect_changepoints()	Identifying episodes with unexpected pivots
Episode encoding (encode_entity_id)	Mapping episode:step → u64 entity_id

Experimental Results

Three benchmarks validate the episodic trace memory hypothesis:

E1: Code Generation (HumanEval) — Results

Static retrieval on 384 MBPP episodes. Finding: CVX ≈ FlatCosine ≈ Random (all ~77% pass@1). For code generation, few-shot formatting matters more than retrieval quality. Static retrieval does not exercise CVX’s temporal features.

E2: Embodied Planning (ALFWorld) — Results

Static plan generation from 336 expert trajectories. Finding: Semantic retrieval >> Random >> Zero-shot (p<0.0001), but CVX ≈ FlatCosine. Without an interactive loop, the temporal structure adds no value over plain cosine search.

E3: Interactive Agent (ALFWorld) — Results

The key result. Real environment execution with step-by-step CVX queries.

Condition	Task Completion
NoMemory	1/30 = 3.3%
CVX-Causal	6/30 = 20.0%

6x improvement when the query is the agent’s actual environment state (not a task description). CVX finds similar mid-episode states in expert memory and returns what happened next — something a flat store cannot do because it has no step ordering or episode structure.

Key Insight

CVX’s temporal features (episode encoding, step ordering, continuation extraction) add value only in interactive settings where the query evolves at each step. Static retrieval reduces CVX to a vector store, where it matches numpy brute-force. The interactive agent loop is the natural deployment pattern for episodic trace memory.

References

• Fang et al. (2025). Memp: Exploring Agent Procedural Memory. arXiv:2508.06433
• Maharana et al. (2024). LoCoMo: Evaluating Very Long-Term Conversational Memory. ACL 2024
• Xu et al. (2025). A-MEM: Agentic Memory for LLM Agents. arXiv:2502.12110
• Position paper (2025). Episodic Memory is the Missing Piece. arXiv:2502.06975

See the full implementation guide for MCP tool definitions and agentic workflow patterns.