001
Every session starts from zero
Context dies when the process ends. The next run re-reads the same files and re-asks the same questions.
Memory infrastructure for AI agents
Persistent episodic, semantic, and procedural memory over MCP, a REST SDK, or the CLI — recall in under 100ms.
The problem
001
Context dies when the process ends. The next run re-reads the same files and re-asks the same questions.
002
Dumping the repo into the prompt is slow, expensive, and still misses the decision you made last Tuesday.
003
The same bug is solved twice because the workaround never became a memory the harness could recall.
Introducing Memoria
Remember on purpose, or emit harness events and let extraction decide. Recall is scored by similarity, recency, and importance, then truncated to a token budget.
session · mcp
$ connected · org_id scoped
✓ remember() semantic 18ms
“prefer pytest over unittest”
✓ recall(“test runner”) 41ms
hit 0.91 · we prefer pytest
▍_
Latency
<100ms
p95 cached recall target
MCP
Stateless
Horizontally scalable adapter
Types
3
Episodic / semantic / procedural
Isolation
100%
Tenant-scoped at query time
Capabilities
[ 01 ]
Similarity, recency decay, and importance — then truncated to a token budget.
[ 02 ]
Near-duplicates reinforce instead of cloning. Sessions merge into one memory.
[ 03 ]
org_id is required on every query. Cross-tenant recall is a bug, not a feature.
[ 04 ]
Harnesses use MCP. Scripts use the Python or Node SDK or the CLI. Same remember / recall / emit contract.
[ 05 ]
Noisy tool calls never hit the queue. Preferences and fixes become memories.
[ 06 ]
Humans sign in with Google. Machines use mem_… keys hashed at rest.
Works with
FAQ