跪拜 Guibai
← All articles
Frontend · Backend · Artificial Intelligence

Agent Long-Term Memory Is a Governance Problem, Not a Retrieval Problem

By zhangxingchao ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most agent memory efforts stop at vector search over chat history, which turns into a junk drawer of raw events that agents cannot reliably act on. A governance-first design — write, revise, forget, promote, scope — is what separates a memory system that actually changes agent behavior from one that just fills the prompt with noise.

Summary

A travel-planning agent that repeats the same scheduling mistake across sessions exposes the real gap: raw conversation logs and RAG-based retrieval are not long-term memory. Without a system to extract, revise, and promote experiences into stable rules, each new task starts from scratch.

The architecture that fixes this splits memory into four layers: raw episodic events, consolidated session summaries, semantic policy rules, and a dynamically built working memory. Candidate experiences are promoted to stable rules only when backed by evidence — cross-session recurrence, tool failures, or explicit user directives — and a background heartbeat handles tidying outside the request path.

Scope isolation, evidence traceability, and rollback capability prevent automatic memory from becoming automatic pollution. The final acceptance test is not whether a rule was recalled, but whether the agent’s output behavior changes in an explainable way when memory is present.

Takeaways
RAG covers only retrieval; long-term memory also requires ingestion, revision, and forgetting to decide what should influence future behavior.
A four-layer model — Raw Episodic, Consolidated Episodic, Semantic/Policy, and Working Memory — separates evidence storage from rule application.
Raw events are evidence sources, not memory; reusable rules like “distant attractions need a full day” must be extracted from failures and corrections.
Candidate memories should carry evidence (original event IDs) so they can be verified and rolled back.
Promotion to stable rules requires cross-session recurrence, tool-failure evidence, or explicit user long-term instructions.
The Semantic/Policy layer must stay small, high-confidence, and traceable to avoid over-generalizing one-off preferences.
A background heartbeat handles memory tidying asynchronously so it doesn’t add latency to user-facing turns.
Un-governed background writes risk turning automatic tidying into automatic pollution from untrusted sources.
New tasks should receive a curated Working Memory brief, not a dump of all historical sessions.
Every memory needs scope tags (project, user, domain) to prevent cross-contamination between agents.
Vector databases serve as semantic recall accelerators; the source of truth must remain in structured, auditable storage.
The real test of long-term memory is whether the agent’s output behavior changes in an explainable way, not whether a rule appeared in the recall list.
Conclusions

Framing long-term memory as a governance pipeline rather than a storage problem shifts the engineering focus from “how do we store more” to “what deserves to influence future decisions.”

The promotion step — moving a candidate memory to a stable rule only when backed by evidence — is the linchpin. Without it, every transient preference risks becoming permanent policy.

Background heartbeat processing is a double-edged sword: it keeps the interaction loop fast but creates a new attack surface where untrusted content can silently poison long-term memory.

Scope isolation is under-discussed in agent memory. A rule learned in a reimbursement context leaking into a travel planner is a failure mode that pure vector similarity cannot prevent.

The behavioral-difference acceptance test is a practical alternative to precision/recall metrics that don’t capture whether memory actually changes what the agent does.

Concepts & terms
Four-Layer Memory Model
A long-term memory architecture splitting storage into Raw Episodic (original events), Consolidated Episodic (session summaries and candidate experiences), Semantic/Policy (stable, high-confidence rules), and Working Memory (a task-specific brief built at runtime).
Memory Promotion
The process of elevating a candidate memory to a stable Semantic/Policy rule, gated by evidence such as cross-session recurrence, tool-failure signals, or explicit user long-term instructions.
Heartbeat Tidying
A background processing loop that periodically scans new raw events, triggers summarization and rule promotion, and writes checkpoints, keeping memory maintenance out of the user-facing request path.
Scope Isolation
Tagging each memory with identifiers like project_id, user_id, and domain so that rules learned in one context (e.g., expense reimbursement) do not leak into unrelated agents (e.g., travel planning).
Working Memory
A dynamically constructed, task-specific context brief — containing relevant stable rules, session summaries, and evidence sources — that replaces the naive approach of dumping all history into the prompt.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗