The State of Agentic Memory
This section explores the fundamental tension in modern AI infrastructure: context management. As we push more data into LLM context windows, we face a tradeoff between retaining critical information and maintaining high reasoning quality. Interact with the slider below to see how the context window dilemma manifests in production workloads.
The Context Window Dilemma
🧠Agent Status
Balanced Processing
Optimal retrieval. The agent has enough context to operate without losing focus.
Reasoning Quality
High
Models degrade when flooded with irrelevant data. Agentic Memory solves this by keeping state out of the prompt until it is explicitly needed.
4-Stage Hybrid GraphRAG Pipeline
Agentic Memory does not just dump text into a vector database. This section breaks down our opinionated Temporal Knowledge Graph architecture. Click through the tabs to understand how we transform raw multi-modal inputs into a structured, queryable, and time-aware intelligence layer.
Chunking & ID
Content is cleanly chunked. Deterministic hashing ensures re-ingesting the same file or conversation is idempotent.
Extraction
The LLM passes over chunks, extracting structured entities (People, Concepts) and the claims/relationships between them.
Temporal Resolution
Classifies into Facts, Events, Instructions, Tasks. Old contradictory memories are marked superseded, not deleted, creating a version chain.
Graph Upsert
Claims are written to Ladybug (hosted) or a legacy graph store (self-hosted). Raw text and facts are embedded into a vector index asynchronously; temporal validity is mirrored in SpacetimeDB on hosted prod.
Beyond Pure Vector Search
This visualization compares standard vector-based RAG architectures against our Hybrid GraphRAG pipeline. Relying purely on vector similarity often fails for complex, multi-hop queries.
By integrating Ladybug (graph store) with a SpacetimeDB temporal sidecar, Agentic Memory excels at tracking evolving state ("When did we switch to SQLModel?"), multi-hop logic, and exact keyword precision, outperforming standard baselines on complex long-running agent benchmarks.
- ✔️ Avoids token burn on retrieval strategies
- ✔️ Explicit citations from graph traversals
- ✔️ Native temporal supersession handling
What you can build
Agentic Memory acts as the universal memory layer across diverse architectures. Below are the core modules available in the open-source framework, representing practical deployment scenarios from individual coding assistants to team-wide institutional knowledge bases.
Codebase Memory
CodeMemory ingests entire repositories into a graph representation. Agents trace function calls, understand architectural decisions, and retain tribal knowledge that usually lives only in senior engineers' heads.
Deep Web Research
WebMemory gives research agents the ability to crawl, chunk, and extract structured claims from the web, building a factual graph of a domain rather than dumping HTML context.
Continuous Chat
ChatMemory provides standalone agents persistent memory across sessions. Your agent remembers previous debugging steps, personal preferences, and evolving project requirements.
Shared Team Memory
On hosted prod, team memory lives on Agent Memory Labs infrastructure — Ladybug plus a SpacetimeDB sidecar on NVMe. Legacy self-hosted deployments may use Neo4j.
Programmatic Interaction
Hosted users integrate via the agent-memory CLI and MCP OAuth on Agent Memory Labs infrastructure (Ladybug + SpacetimeDB on NVMe). The Python SDK below is a historical self-hosted example; legacy operators may still run Neo4j, SpacetimeDB, or local SQLite.
Historical Python SDK (legacy self-hosted)