Research basis¶
coniti’s design draws on memory research and on existing agent-memory systems. These sources inform the roadmap and the shape of the framework; they are not runtime dependencies.
How these were read
For the papers below only the abstracts were read, not the full texts. The summaries are paraphrases, and claims about them should be checked against the papers.
Memory theory¶
- Complementary learning systems (McClelland et al., 1995): fast episodic capture alongside slow consolidation.
- Spreading activation (Collins and Loftus, 1975): retrieval as activation spreading through associations.
Retrieval and structure¶
- RAPTOR: tree-organised retrieval.
- GraphRAG: graph and community summaries.
- HippoRAG: hippocampal-index style memory.
- Zep: temporal graph memory.
- A-MEM: Zettelkasten-style linking of notes.
Agent memory¶
- MemGPT: tiered memory.
- CoALA: a taxonomy of agent memory.
- Generative Agents: observe, reflect, retrieve.
Learned memory¶
- Memory-R1: learned memory operations via reinforcement learning.
- Reflexion: feedback-driven improvement.
Design inspiration for the memory backend and dependency injection¶
These shaped the design. None is a dependency of coniti.
- LangGraph checkpointing, overview and the InMemorySaver reference: the split between a saver interface and an in-memory implementation behind it inspired
MemoryBackendandInMemoryBackend. - FastDepends: the
Depends(...)marker style for declaring collaborators. - dependency-injector: overriding: the idea of a temporary override that restores earlier bindings.
Note
The random-projection technique used by the random_projection embedder is a
standard one; its rationale was not independently sourced here.