Types of Agent Memory¶

Just like humans, agents need to remember things to be effective. Below is a concise list of the key memory types involved in building reliable AI agents, without grouping into primary/secondary.

System and policy prompt¶

Persistent instructions and policies that steer behavior (role, capabilities, safety and business rules). This is not Cache Augmented Generation (CAG).

Conversation history (episodic)¶

Message transcript (often summarized) that preserves context and coherence across turns.

Working memory / scratchpad¶

Internal reasoning artifacts (plans, intermediate computations). Sometimes persisted for traceability.

Tool and capability memory¶

• Tool request/response transcript included in context for grounding.
• Short-lived tool state (pagination cursors, partial results).
• Capability schemas/affordances; see 3-capabilities.md.

Goal and plan state¶

Current objectives, subgoals, and plan-of-action, whether explicit (state machine/flow engine) or implicit (prompted plan).

Knowledge memory (RAG)¶

External knowledge indexed for retrieval (documents, code, KBs).

• Index types: vector, keyword/lexical, and graph (often combined).
• Consider freshness, provenance, and retrieval accuracy.
• See 2-accessing-knowledge.md and 5-document-store.md.

User profile and preferences¶

Identity, roles, preferences, and constraints for personalization. Requires consent, retention controls, and audit.

Self-improving profile memory¶

A governed mechanism to learn and update user traits over time.

• Signals: explicit inputs, in-chat confirmations, choice history, engagement signals, and tool outcomes.
• Representation: typed key–value traits with confidence (0–1), evidence pointers, and last_seen_at timestamps.
• Updating: online updates with decay; add traits only after a confidence threshold; resolve conflicts by lowering confidence and requesting confirmation when needed.
• Personalization: include only high-confidence traits in prompts or retrieval filters; prefer structured context over free text.
• Controls: user consent, view/edit/reset UI, per-field TTL/retention, scope (per-user, per-agent, or org), and full audit trail.
• Safety: avoid sensitive/protected attributes unless strictly necessary; minimize PII; encrypt at rest and in transit; consider differential privacy for aggregated learning.

Episodic summaries (past sessions)¶

Summaries of prior interactions and outcomes to maintain continuity without storing full transcripts.

Organizational/shared memory¶

Team- or org-level facts, standards, and decisions accessible to multiple agents.

Safety and policy memory¶

Compliance constraints and allow/deny lists, maintained separately from prompts for auditable updates.

Model-embedded (parametric) memory¶

What the model “knows” from pretraining or fine-tuning (including adapters/LoRA).

• Pros: fast and compact at inference.
• Cons: hard to update; may hallucinate without external grounding.

Operational memory and caches¶

Mechanisms that accelerate or stabilize behavior.

• Caches (Cache Augmented Generation, CAG): prompt/output and embedding caches with invalidation to avoid staleness.
• Transformer KV cache: ephemeral per generation to speed decoding.
• Checkpoints and flow state: persisted progress for long-running or resumable processes.

Governance, privacy, and security¶

Cross-cutting concerns for any memory that contains user or sensitive data.

• Retention/TTL and redaction; per-field sensitivity classes.
• Encryption in transit and at rest; see ../n-encyclopedia/chat-message-encryption.md.
• Access control, consent, and purpose limitation.
• Auditing and observability; see ../n-encyclopedia/logging.md.

Quick reference: dimensions¶

• Parametric vs external: in-model vs in a store (RAG, profiles, summaries).
• Ephemeral vs persistent: within one run/session vs across sessions.
• Private vs shared: per-user/agent vs organizational/multi-agent.
• Grounded vs generative: retrieved facts vs model’s internal knowledge/synthesis.