Overview contents
Research question
How can useful memory be built across sessions without turning incorrect, outdated or unauthorised information into seemingly reliable context?
Technical context
A language model does not automatically retain a permanent, reliable memory of interactions. Larger context windows allow more material to be supplied, but do not solve selection, continuity, conflicts, permissions or deletion.
Persistent memory must determine which information to record, how to structure it, when to update it and what to retrieve. It is therefore more than a simple vector database: it must handle identity, time, versions, provenance and permissions.
What we are studying
- 01
Relational data for facts, versions, identities and permissions.
- 02
Document archives and citations to preserve the link with sources.
- 03
Semantic and lexical search, with graphs where needed for complex relationships.
- 04
Rules for writing, updating, conflicts, abstention and deletion.
Applications to explore
- CRM with customer history, decisions and next actions.
- Organisational memory for procedures, meetings and documents.
- Multi-user assistants with granular permissions and inspectable sources.
- Operational agents that retain the state of processes and open tasks.
How it can combine with other technologies
This architecture could integrate with CRM systems, document archives, chat and operational applications, and work with models from different providers. Citations, version management and permissions can make the provenance and scope of retrieved information verifiable.
What a prototype needs to establish
- V01
Accuracy and coverage of retrieved information.
- V02
Handling time, updates, conflicts and abstention.
- V03
Effective isolation between users, organisations and projects.
- V04
Traceability, correction and deletion of memories.
- V05
Resistance to malicious content, latency and operating cost.
Known limitations and risks
Incorrect or outdated memory can make an answer more convincing but less reliable.
Risks include data leakage, incorrectly applied permissions, prompt injection from sources, profiling and incomplete deletion.
Technical references