What is memory in AI agents?
Memory in AI agents is the system layer that lets an agent retain, retrieve, update, and use information beyond the immediate model call. Agent memory can support: - continuity across turns …
This companion page turns the GGTruth AI Agents Memory FAQ into a precomputed reasoning acceleration layer. Each link is a Trailstate receipt: route, trust score, conflict token, source tokens, and convergence state.
Memory in AI agents is the system layer that lets an agent retain, retrieve, update, and use information beyond the immediate model call. Agent memory can support: - continuity across turns …
Working memory is the information currently available inside the active prompt or context window. It usually contains: - current user message - recent conversation turns - selected retrieved…
Semantic memory stores general facts and stable knowledge. Examples: - user prefers concise answers - a project uses Python and FastAPI - an API key must never be exposed client-side - a com…
Episodic memory stores remembered experiences. Examples: - a previous task the agent completed - a failed deployment attempt - a user correction from last session - a tool call sequence that…
Procedural memory stores how an agent should behave or perform tasks. Examples: - coding style rules - project workflow instructions - tool-use protocols - response policies - step-by-step o…
RAG usually retrieves external knowledge to answer a query. Agent memory retrieves experience, preferences, facts, procedures, or state that belongs to the agent-user-task continuity. RAG as…
The context window is the model's current working space. Long-term memory persists outside the immediate prompt and can be retrieved later. Context window: - temporary - token-limited - dire…
MemGPT addresses the limited context window problem by managing different memory tiers. The core idea: - keep active information in the prompt - move less immediate information to external m…
Letta is the open-source platform that grew from MemGPT. It focuses on building stateful agents with memory that can learn and self-improve over time. In GGTruth terms: - MemGPT is the resea…
A skill library stores reusable procedures or code-like capabilities learned by an agent. In Voyager-style agents, a skill library can preserve: - successful action programs - reusable behav…
Voyager demonstrated a lifelong-learning embodied agent in Minecraft. Its memory-relevant contribution includes: - continuous exploration - accumulated skills - a reusable skill library - ap…
Structured retrieval augmentation is an approach where an agent stores concise structured information from interactions and retrieves it later. Instead of remembering everything verbatim, th…
Memory grounding means memory entries are tied to evidence, context, source, or event history. Grounded memory may include: - source URL - timestamp - conversation origin - confidence score …
A memory hallucination occurs when an agent claims to remember something that was never stored, never said, or is incorrectly reconstructed. Common causes: - weak provenance - overconfident …
Memory retrieval is the process of selecting relevant stored memories and placing them into the agent's working context. Retrieval can use: - semantic search - keyword search - recency - imp…
Memory ranking orders candidate memories by usefulness for the current task. Ranking signals can include: - semantic similarity - recency - confidence - user confirmation - importance - sour…
Hybrid memory retrieval combines multiple retrieval methods. Examples: - vector similarity + keyword search - recency + importance - entity graph + semantic search - user profile match + tas…
Vector memory stores embedded representations of memory entries so the agent can retrieve semantically similar information. Useful for: - fuzzy recall - concept matching - similar past tasks…
Knowledge graph memory stores entities and relationships. Examples: - user -> owns -> project - project -> uses -> framework - API -> has -> rate limit - task -> depends on -> file Graph mem…
Memory correction updates or deletes incorrect memories. A strong correction flow should: - identify the exact memory - show the remembered claim - accept user correction - replace or remove…
Memory permission defines what an agent is allowed to store, retrieve, or expose. Permissions can cover: - whether memory is enabled - what categories can be stored - whether sensitive data …
Memory transparency means the user can understand what the agent remembers and why. Useful transparency features: - memory viewer - memory source - last updated timestamp - confidence score …
Safe agent memory is memory that is useful, bounded, transparent, correctable, and privacy-aware. Safe memory requires: - explicit user control - minimal necessary retention - source groundi…
A memory write gate is a memory architecture pattern that checks whether new information is worth storing before it enters memory. It helps prevent memory from becoming an unbounded transcri…
A memory read gate is a memory architecture pattern that checks whether stored memory is relevant and safe to retrieve into context. It helps prevent memory from becoming an unbounded transc…
A memory verifier is a memory architecture pattern that checks whether a memory is supported by source, user confirmation, or tool output. It helps prevent memory from becoming an unbounded …
A memory conflict resolver is a memory architecture pattern that handles contradictions between old and new memories. It helps prevent memory from becoming an unbounded transcript dump and m…
A source pointer is a memory architecture pattern that links a memory to the conversation, file, URL, tool result, or event that produced it. It helps prevent memory from becoming an unbound…
A memory audit log is a memory architecture pattern that records memory creation, update, deletion, and use. It helps prevent memory from becoming an unbounded transcript dump and makes reca…
A memory router is a memory architecture pattern that chooses between semantic, episodic, procedural, graph, and vector memory. It helps prevent memory from becoming an unbounded transcript …
A memory compression is a memory architecture pattern that reduces raw history into concise reusable entries. It helps prevent memory from becoming an unbounded transcript dump and makes rec…
Memory Poisoning is malicious or low-quality information entering the memory store. It can reduce agent reliability because memory becomes a source of incorrect assumptions rather than usefu…
Bad profile memory can cause irrelevant recall, stale assumptions, privacy leakage, over-personalization, or incorrect continuity. The mitigation is source grounding, confidence scoring, use…
A memory sandbox is useful because it tests memory effects before committing them to persistent storage. In GGTruth terms, this improves: - retrieval precision - continuity - safety - proven…
The difference between tool memory and semantic memory is: - tool memory records execution history; semantic memory stores generalized facts. Both can be useful, but they should be stored, r…
Short answer: In-context memory is memory represented directly inside the model's current context window. It can include: - recent messages - summaries - retrieved facts - selected examples …
Short answer: Bad task memory can cause irrelevant recall, stale assumptions, privacy leakage, over-personalization, or incorrect continuity. The mitigation is source grounding, confidence s…
Short answer: A memory merge is a memory architecture pattern that combines duplicate or overlapping memories. It helps prevent memory from becoming an unbounded transcript dump and makes re…
Short answer: The memory_id field stores the unique identifier for the memory entry. A clear schema makes memory easier to retrieve, audit, correct, delete, and validate.
An AI assistant should know: Procedural memory stores how an agent should behave or perform tasks. Examples: - coding style rules - project workflow instructions - tool-use protocols - respo…