RESEARCH

Memory Survey (Dec 2025)

Unified taxonomy of agent memory: forms, functions, and dynamics.

Memory Survey: A Unified Taxonomy (Dec 2025)

A comprehensive survey paper from December 2025 proposed a unified taxonomy for agent memory systems, organizing the landscape along three dimensions: forms, functions, and dynamics.

Three Dimensions

1. Memory Forms (What is stored)

  • Parametric Memory - Knowledge encoded in model weights during training.
  • Textual Memory - Natural language stored in external systems (our primary focus).
  • Embedding Memory - Dense vector representations for semantic search.
  • Structured Memory - Knowledge graphs, databases, and tables.

2. Memory Functions (How it is used)

  • Episodic - Event sequences and experiences (session logs, conversation history).
  • Semantic - Facts, concepts, and general knowledge (extracted insights).
  • Procedural - Skills, workflows, and tool-use patterns (how-to knowledge).

3. Memory Dynamics (How it changes)

  • Accumulation - Adding new memories over time.
  • Consolidation - Merging, summarizing, and compressing memories.
  • Forgetting - Decay, pruning, and active deletion of stale memories.
  • Retrieval-induced modification - Memories change when accessed (confidence updates, access counts).

Key Findings

  1. Most systems only implement accumulation - Few handle consolidation or forgetting well.
  2. Retrieval quality matters more than storage capacity - Better retrieval beats more memory.
  3. Memory safety is under-studied - Poisoning, hallucination reinforcement, and privacy are open problems.
  4. Evaluation is inconsistent - No standard benchmarks across all memory types.

Design Principles from the Survey

  1. Treat memory as a runtime policy problem, not just a storage problem.
  2. Budget-aware routing: choose retrieval complexity per query.
  3. Hierarchical retrieval: theme -> episode -> raw evidence.
  4. Active forgetting: use decay and reinforcement to control growth.
  5. Memory safety: treat poisoning and self-reinforcing bad memory as threats.

Relevance to Memory Platform

The taxonomy maps directly to our system:

Survey ConceptPlatform Feature
Textual + Structured MemoryPostgres memories with JSON context
Episodic FunctionSessions + session-scoped memories
Semantic FunctionProject memories with categories
Procedural FunctionPattern/procedure category memories
AccumulationMemory creation via API/MCP
ConsolidationMemory evolution (merge/compress)
ForgettingConfidence decay + access tracking

References

  • "A Survey on Memory Systems for LLM-based Agents" (Dec 2025)
  • The survey covers 100+ papers from 2023-2025
  • Proposed evaluation framework for memory systems