memomee
Memory control plane · v0.4

Agents need more than memory.
They need continuity.

Memomee is the memory governance and retrieval control plane for long-running AI agents. Recall broadly, govern rigorously, inject precisely — task state, decisions, and evolving facts kept compact, trustworthy, and resumable across sessions.

p99 retrieve
< 100 ms
events
append-only
facts
versioned · superseded
live tasktask_8f3a · refactor-auth-v2
GET /v1/retrieve42 ms
T+0
·
T+12m
·
T+38m
T+1h
f
T+2h
!
T+2h
·
T+3h
now
checkpointeventfactwarning
supersedes:f-29
retrieve_bundle.jsonlive · trace included
{
"checkpoint": { snapshot_id: "snap_20260509_018", stage: "verify" },
"events": [ 12 returned, // 34 filtered ],
"facts": {
"active": [ "budget_approved → $80,000" ],
"superseded_summary": [ "$50,000 → $80,000" ],
},
"warnings": { contradictions: 1 },
"trace": { ranking: "recency + confidence", query_latency_ms: 42 },
}
The real memory problem

Most agents don't fail loudly.
They fail by losing the plot.

An agent starts strong. Then a workflow branches. A session ends. A fact changes. A summary compresses away a key decision. Now the agent is still running — but no longer grounded.

Summary drift

Rolling summaries quietly compress away the decision that actually mattered.

Episodic log

Retrieval mismatch

Similarity search finds something relevant — to a different problem.

Explainable retrieval

Stale facts

Yesterday's truth keeps shaping today's decisions, silently.

Versioned facts

Workflow loss

An interruption ends a session and the agent has to rediscover everything.

Checkpoints

Context fragmentation

State splits across threads, tools and providers; teams glue it back by hand.

Task-scoped bundle

The long version: why memory fails — five chapters on how agents lose the plot.

PILLAR 01 / 3
Govern rigorously · checkpointed working state

Resume exactly where the work left off.

Capture the live state of a task — stage, goals, parameters, the last successful step — so an interruption doesn't force a restart.

  • Immutable revisions; every checkpoint kept for replay
  • Opaque resume tokens map back without leaking state
  • Snapshot ids let external systems reference a moment in time
checkpoint.py
PUT /v1/checkpoints
# resume tokens are opaque, not state
client.checkpoints.update(
  task_id="task_8f3a",
  workflow_stage="verify",
  active_goals=["verify auth flow", "fix regression"],
  last_successful_step="run_unit_tests",
  resume_token="tok_abc123",
)

# → snapshot_id: snap_20260509_018

The vault in depth: the event ledger.

event.py
POST /v1/events
client.events.write(
  task_id="task_8f3a",
  actor="coding-agent",
  event_type="decision",
  payload_summary="branched to fix-regression sub-task",
  source_metadata={
    "model": "claude-4",
    "tool": "test_runner",
  },
)
project_alpha · budget_approved 3 versions
Apr 24
finance_system · v1
$25,000
superseded
supersedes ↓
Apr 30
finance_system · v2
$50,000
superseded
supersedes ↓
May 9
finance_system · v3
$80,000
active
PILLAR 02 / 3
Govern rigorously · append-only ledger + versioned facts

Update what's true without silently breaking trust.

A write-once event ledger holds what actually happened — actor, type, payload, source. Facts evolve through explicit supersession, not overwrite. Every version keeps its provenance and link — so retrieval can show what changed and why.

  • Events are immutable; trigger-enforced append-only
  • active / superseded / retracted as first-class status
  • Lineage chains render the full history of any (entity, relation)
  • Confidence and source travel with every fact

The engine in depth: the governance model.

PILLAR 03 / 3
Recall broadly, inject precisely · explainable retrieval

See what was surfaced — and what wasn't.

Every retrieval carries a trace: returned vs filtered counts, ranking rationale, recency and source signals, contradictions. Consolidation merges repetitive events without losing the evidence — every output keeps a reference back to the raw events.

  • Bundle separates checkpoint, events, facts, warnings, trace
  • Idempotent consolidation — same window in, same output out
  • Contradictions surfaced as warnings, not silently merged
  • Lineage links from compacted summaries back to event_ids
retrieve_bundle · trace req_b91c · 42 ms
events
12 /46
facts
8 /20
checkpoint
1 /1
ranking   recency + confidence
recency   most recent event 2026-05-09T14:32Z
sources   user_input(5) · tool_output(4) · system(3)
warnings   contradictions: 1
contradiction_warning.json
POST /v1/consolidate
{
  "events_compacted": 12,
  "facts_promoted": 3,
  "contradictions_found": 1,
  "warnings": [{
    "entity": "project_alpha",
    "relation": "budget_approved",
    "fact_ids": ["f-29", "f-44"],
    "resolution_status": "open",
  }],
}

Reflection, measured: the evidence tiers.

How it works

Four steps. One governance layer.

01

Capture

Agents write events and checkpoints as they work. Every action is append-only.

02

Consolidate

Rules-based compaction merges repetitive events and flags contradictions.

03

Retrieve

Bundle returns checkpoint, filtered events, active facts, superseded summary, and trace.

04

Explain

Trace shows what was returned, what was filtered, and why — with contradiction warnings.

live tasktask_8f3a · refactor-auth-v2
GET /v1/retrieve42 ms
T+0
·
T+12m
·
T+38m
T+1h
f
T+2h
!
T+2h
·
T+3h
now
checkpointeventfactwarning
supersedes:f-29

The four steps above are the short version — the full six-stage pipeline, box by box: how it works.

Who it's for

Built for agents that run longer than a session.

Coding agents

Resume a refactor across sessions — repo state, decisions, and open threads survive a restart.

Research agents

Carry findings, sources, and contradictions across a long investigation without re-deriving them.

Enterprise workflow

Long-running approvals and hand-offs that stay resumable and auditable across days.

Support & ops

Per-customer memory with governance — isolation, supersession, and a full audit trail.

Multi-agent handoffs

Pass task state between agents without losing the plot or duplicating work.

Problem → architecture → what changes: the use cases, in depth.

The category

Recall broadly. Govern rigorously.
Inject precisely.

Other products help agents store more context. Memomee holds recall@k parity and leads open-source RAG and mem0 at a matched regime — at a fraction of the tokens — then governs what should survive, what should change, what should be forgotten, and what should shape the next decision.

0.98–1.00
recall@k parity
matched against a full-context reader
3–11×
token discount
fewer tokens read per question, recall held at parity
0 leaks
governed recall
cross-user isolation, abstention, supersession — CI-gated
See the full head-to-head — numbers, methodology, and what we don't claim
Production-grade by default

Built for memory you can defend in a review.

Audit
Append-only events
Write-once, read-many. Trigger-enforced at the database level.
Isolation
Multi-tenant
Org-scoped queries; no endpoint accepts an org_id from requests.
Compliance
GDPR deletion
User-scoped erasure: derived memory deleted with chain repair, events redacted in place, audit-logged.
Encryption
TLS · AES-256
TLS 1.3 in transit; AES-256 at rest on the managed deployment.
Abstention
Refuse on conflict
Prefers explicit uncertainty over unsafe recall — abstains when memory is missing, conflicting, stale, or out of scope.

The enterprise picture: what the system enforces — and what we don't hold yet.

Stop rebuilding context

Start governing it.

Your agents shouldn't have to rediscover the world every time a workflow branches or a session restarts. Memomee is invite-first while we finish the managed product — join the waitlist for early access.

Or email hello@memomee.ai · see the head-to-head