01 / OBSERVE
Record what happened.
The host records the execution, outcome, evidence, and the agent’s interpretation of the result.
execution episodeDurable experience for autonomous agents
Memory that changes the next decision.
Engram turns completed executions into bounded, reusable experience—so a fresh agent session can remember what mattered, act differently, and stay within its authority.
16/16scenarios passed
3agent surfaces
0unauthorized escapes
Session A · observed
Atlas missed the urgent SLA.
Two failures are recorded as experience—not buried in a transcript.
● MEMORY ADMITTED
MEMORYSLICE · ELIGIBLE
Use Beacon for urgent tasks.Bounded claim · provider selection onlyINFLUENCEGRANT · AUTHORIZED
Session B · fresh runtime
Switch to Beacon.
The agent requested memory, cited it, and changed its provider decision.
● OUTCOME UPDATED
PERSIST → RECALL → INFLUENCE → UPDATE
02 · The memory loopexecution → experience → influence → update
Experience becomes a capability the next session can use.
Engram does not replay an entire transcript. It extracts what happened, checks when it applies, and returns only the bounded lesson the agent is allowed to use.
02 / FORM
Form reusable experience.
Engram turns the event into an admitted memory with applicability, confidence, and source lineage.
Sibyl · durable memory03 / INFLUENCE
Return only what applies.
A fresh agent requests recall. Engram filters the memory and limits the effects it can influence.
bounded authority04 / LEARN
v1 → v2
Update from the outcome.
The result is evaluated and a new immutable memory version is created—stronger, weaker, or invalidated.
outcome evaluationDurable memory, not a longer prompt. The model stays the same. Its experience does not.↺ NEXT SESSION · NO RAW HISTORY · NO NEW AUTHORITY
03 · Bounded influencememory changes strategy · never authority
The safety boundary
Remember more. Never assume more power.
Engram lets an agent use experience to choose a better path. The memory arrives with a grant that names exactly what it may influence—and everything it may not.
Atlas → BeaconProvider selection changed. Budget, signer, wallet, and capability stayed exactly where they were.
InfluenceGrant · allowed
✓
Change operational choices.
Memory can guide bounded operational choices inside the current mandate.
- provider selection
- tool selection
- retry policy
- timeout policy
- fallback policy
InfluenceGrant · denied
×
Do not expand power.
Memory cannot create authority that was not already granted to the agent.
- increase budget
- create signer
- create wallet
- create asset
- create capability
AUTHORITY NEVER INCREASES
Experience is reusable. Authority is not transferable.NO RAW HISTORY · FAIL CLOSED · RECEIPT BASED
04 · Agent surfacesone runtime · three ways in
Built for the way agents work
One memory layer. Any agent surface.
Engram keeps the domain and authority model canonical while giving agents the interface that fits their runtime.
Same lifecycle everywhere. Record an execution. Recall applicable experience. Request bounded influence. Evaluate the outcome.
01 / TYPED
{ }
SDK
For application code that wants typed requests, stable errors, and direct lifecycle control.
ENGRAMCLIENT · TYPESCRIPT02 / HTTP
REST
For services and frontends that need a bounded, observable API boundary.
/V1/MEMORIES/RECALL03 / AGENT
MCP
For tool-using agents that request memory and influence as explicit actions.
TOOLS/CALL · STDIOrecord→recall→influence→evaluate
One canonical runtime. No parallel memory semantics.SDK · REST / HTTP · MCP STDIO
05 · Evidence, not promisessee the receipt · run the proof
The result
Memory is useful when it changes what happens next.
One fresh-session proof. One bounded provider decision. A durable memory update you can inspect instead of a claim you have to take on faith.
Fresh-session comparison
Same task. Same model. A0 is the baseline without Engram memory. A2 is the same agent with MCP recall and bounded influence.
Decision
A0 · NO MEMORY
A2 · ENGRAM
Provider selected
Atlas
Beacon
Memory requested
NONE
RECALLED
Influence result
NONE
BEACON AUTHORIZED
Memory result
NO UPDATE
VERSION UPDATED
16/16scenarios passed
3agent surfaces
0unauthorized escapes
LLAMA 3.2 3BPROOF · MCP ACTIVE AGENT
Build agents that remember the lesson. Not just the transcript.LOCAL EVIDENCE · NO LIVE UPLIFT CLAIMED