Agent Memory
Durable, governed memory for AI agents and chats
Lore gives AI conversations durable memory you control. Memorable facts are extracted from a conversation into a memory bank, and the relevant ones are recalled into later prompts at the moment of inference. Each bank has its own models, retention window and off switch, recall is strictly opt-in, and an unreachable memory service never blocks a conversation. Chat and agent runs carry context forward instead of starting cold.
Memory makes AI useful across sessions, and it is also where sensitive context builds up. Lore keeps memory in banks your organization controls, each with its own models, retention, and off switch, encrypted under your organization's keys, recalled only where you opt in, and audited without exposing what was remembered.
- Memory banks with their own models, retention, and kill switch
- Recall on chat, agent runs, API traffic, and tool conversations — opt-in and fail-open
- Memories encrypted under your organization's own keys
- Share banks with people and teams
- Audit what was recalled, never the memory text
What you get
What Lore gives your team.
Memory banks you govern
Each bank has its own embedding and extraction models, retention window and enable switch. Share banks with people, teams and organizations.
Automatic extraction and recall
Facts are extracted after a conversation turn, duplicates are stored once, and the most relevant memories are recalled before the next model call.
Opt-in and fail-open
Memory only engages where you assign it, and a missing bank or unreachable service never blocks a conversation.
Memory for agents too
Connect an agent to a bank by name, choose recall, extraction or both, and see which agents use which banks.
Audit without exposing memory
A recall history shows what was recalled, when and by which model, never the memory text. Reading memory text is a separate, all-or-nothing request.
Encrypted, retained, recoverable
Memories are sealed under your organization's own keys, expire on per-bank retention windows, and deleted banks and memories are recoverable from the recycle bin.
How it works
- 1
Create a memory bank
Make a bank, picking its embedding and extraction models by name and setting a retention window. Share it with people, teams or organizations, tag it, and document it.
- 2
Assign it where memory should apply
Lore is opt-in: assign the bank through the AI-path guard assignment to users, teams, pools or sources. Bind agents to a bank by name, choosing recall, extraction or both.
- 3
Let conversations build and recall
After each turn, memorable facts are extracted and duplicates stored once. Before the next model call, the most relevant memories are recalled, and replies show how many informed them.
- 4
Audit, re-embed and retire
Browse memories and recall history without exposing question text, review activity on the dashboard, re-embed after a model swap, and retire banks through the recycle bin.
Ask AI about Lore
Every Lore page has an Ask AI button. It opens a chat that already has the page, its rows, and your selection. Try asking:
- Which memory banks here are disabled, and what retention window does each have?
- Which bank had the most failed extractions over the last 30 days?
- Which agents are bound to this bank, and do they recall, extract or both?
How teams use it
- Let a support assistant remember customer preferences across separate chats.
- Give an operations agent standing context about your environment without re-explaining it each run.
- Switch memory off for one team's bank instantly without touching other banks.