Skip to content

Human-verified memory for AI agents

Your tools capture everything.SPONGE turns it into trusted memory.

SPONGE sits after meetings, transcripts, CRM, email, and documents. It turns raw records into candidate memories, gives people the evidence to review and approve them, then lets every agent query the same source-backed record.

  • transcript · Mar 2“Renewal is March 14.”
  • transcript · Mar 2duplicate“Renewal is March 14.”
  • email · Apr 1“Actually — renewal moved to April 14.”
  • crm note · May 8Prefers email over phone.
  • chat · Jun 12uncertain“Might move to annual billing?”
memory acme-corp
Renewal dateApril 14 human approvedcurrentsupersedes Mar 143 sources
Preferred contactEmailapprovedai confidence 0.94
Billing planAnnual — unconfirmedwithheld · awaiting review
human approved · just now
raw records in · approved memory out

The problem

More context does not create reliable memory.

Transcripts and records contain duplicates, tentative ideas, corrections, contradictions, and facts that stopped being current. Retrieval can find every version, but it cannot decide which version your organization has approved for an agent to use.

Raw history

Mar 14 — contract renewaltranscript
Mar 14 — contract renewalcopy
Apr 14 — renewal changedemail
“maybe moving renewal”chat
“customer said April”slack

Five records. Three versions of one fact.

Retrieval alone

0.92Mar 14 — contract renewal
0.91Apr 14 — renewal changed
0.90Mar 14 — contract renewal
0.86“customer said April”

Four close matches, equally weighted. The agent still has to guess.

Human-approved memory

Contract renewalApril 14 human approvedcurrentsupersedes Mar 143 sources
approved · 2 min ago

One approved current answer, with its evidence and history attached.

How it works

From raw records to approved memory.

  1. 01

    Absorb

    Connect or import the places knowledge already lives — meeting transcripts, CRM records, email, documents, and agent observations — without replacing the tools that captured them.

  2. 02

    Reconcile

    AI turns raw records into candidate memories — not verified truth. It groups duplicates, detects corrections, surfaces contradictions, and attaches the supporting evidence to each candidate.

  3. 03

    Review & approve

    A person reviews each proposed memory alongside its meeting, speaker, exact quote, AI confidence, and history — then edits, approves, rejects, merges, or marks it temporary. Only approved items enter verified memory.

  4. 04

    Query

    Agents query approved memory by default and get the current answer with its provenance, approval state, and revision history — no reinterpreting the raw transcript every time.

The layer after capture

SPONGE is not another meeting notetaker.

Capture tools tell you what was said. Retrieval systems find relevant passages. SPONGE is the control layer that turns those records into memory people have approved and agents can safely reuse.

Capture and retrieval tools

StoresTranscripts, summaries, documents, and chunks
Optimizes forCapture and relevance
Can answer

“What was said, or looks relevant?”

SPONGE

MaintainsApproved memory + evidence + state
Optimizes forIntegrity, control, and accountability
Can answer
  • What has been approved as current?
  • Who approved it, and from which evidence?
  • What did it replace, and when?
  • What remains tentative, conflicting, or temporary?

AI proposes. People decide. Agents retrieve.

The verified memory object

Every approved memory carries its proof.

SPONGE’s unit is not a text chunk. It is a human-approved claim connected to its source evidence, current state, approval history, relationships, and the memory it superseded.

  • Approval and AI confidence are separate signals. A model’s score never stands in for a person’s decision.
  • Every approved value links to its exact source — the meeting, the speaker, the quote, and the timestamp.
  • New approvals replace without erasing. The current value changes; the previous one stays in the revision history.
  • Temporary, disputed, or unapproved claims stay outside verified memory until someone decides otherwise.
Conceptual interface
memory mem_9f2c4 · acme-corp
Customer pricing preferenceAnnual contract preferred human approvedcurrent
AI confidence0.96
ApprovedAug 19
Approved byWorkspace owner
Revisionv2 of 2
Sources · 3
callM. Chen — “Annual works better for our budgeting.”Feb 12
emailProcurement — “Confirming annual terms.”Jun 3
callM. Chen — “Yes, keep us on annual.”Aug 19
Supersedes · revision 1
Month-to-month preferenceFeb 12 – Jun 3 · kept in history
Related
Acme CorpProcurementContract negotiation
approved · Aug 19

For agents

Agents read from memory people have approved.

SPONGE gives ChatGPT, Claude, custom agents, and internal copilots one model-independent memory layer. Agents query approved memory by default. New observations can create proposals, but they do not silently change what is verified.

ConversationsCRMEmailDocsTool output
SPONGE
ai reconciliation → candidate memories, with evidence
Human review & approvaledit · approve · reject · merge · mark temporary
Verified memoryapproved · source-backed · versioned
ChatGPTClaudeGeminiCustom agentInternal copilot
The model can change. The approved memory and audit trail remain.

For developers

Query approved memory, not raw context.

Ask a question and receive the approved current answer with its status, reviewer, source evidence, and revision history — instead of a pile of similar transcript chunks.

SPONGE is pre-release. The interface shown here is the target developer experience — the API surface we’re designing toward — published for feedback, not yet a shipping SDK or schema.

resolve.tsTarget developer experience
Request
const memory = await sponge.memory.resolve({
  subject: "acme-corp",
  question: "What renewal date should we use?",
  status: "approved",
})
Response
{
  "value": "2027-04-14",
  "status": "current",
  "verification": "human_approved",
  "approved_by": "revops-lead",
  "approved_at": "2027-04-02",
  "supersedes": "2027-03-14",
  "revisions": "mem_9f2c4/history",
  "sources": [{
    "meeting": "mtg_0412",
    "speaker": "M. Chen",
    "quote": "Let’s move renewal to April 14.",
    "at": "2027-04-01T16:04Z"
  }]
}

Product architecture

A human approval boundary between raw records and AI action.

SPONGE separates source evidence, candidate memory, and verified memory. AI can extract and reconcile. People decide what becomes durable. Agents consume only the approved layer by default.

Evidence stays linked

Every candidate and every approved memory keeps its connection to the meeting, speaker, quote, and timestamp it came from.

Approval is explicit

Nothing becomes verified memory implicitly. A person reviews, and the record shows who approved what, and when.

History remains intact

New approvals supersede older values without deleting them. The revision history is part of the memory, not a log somewhere else.

Conceptual product architecture

Sources

Meeting transcripts · CRM · email · documents · tool outputThe records your existing tools already capture.

Candidate-memory pipeline

ExtractionCandidate facts, decisions, commitments, and changes.
Entity resolutionThe same customer, project, or person across records.
Duplicate · contradiction · change detectionGrouped and flagged, with evidence attached.

Human approval boundary

Review, edit, approve, reject, merge, or mark temporaryThe evidence sits alongside every decision. Only approved items pass.

Verified-memory layer

Approved current stateWhat your organization has accepted as current.
Source evidence & approval metadataWho said it, who approved it, and when.
Relationships & revision historyWhat connects to what, and what superseded what.

Access layer

API · search · agent retrievalHow approved memory is read; future connectors planned.

Consumers

Assistants · copilots · workflows · custom agentsThey query the approved layer by default — and can propose, not overwrite.

Conceptual product architecture. SPONGE is being built toward this workflow; not every integration or interface shown is currently available.

Why now

Agents are taking on long-term responsibility. Memory needs governance.

The model can change. Approved memory should remain.

Starting with meeting-driven work

Turn conversations into memory every agent can use.

Revenue teams

Preserve current customer commitments, preferences, pricing, renewal details, and changes across meetings and connected records.

Operations & project teams

Maintain approved decisions, action items, owners, risks, and deadlines.

AI assistants

Carry approved personal or company context across sessions without replaying raw history.

Agent builders

Add human-governed, source-backed memory without building the full review and reconciliation system from scratch.

Trust

Human approval is the trust boundary.

SPONGE can organize evidence and recommend what changed. It does not silently promote a model inference into organizational truth.

  • AI proposes; people approve.
  • Every approved memory links to its exact evidence.
  • Approval records who reviewed or changed the memory.
  • Corrections update the current state without deleting history.
  • Temporary and uncertain information stays outside verified memory until approved.
  • Users can inspect, update, export, and delete their memory.

Give your agents memory your team can stand behind.

SPONGE is building the human-verified memory layer between your records and the AI systems acting on them.

Or email us directly at hello@spongememory.com.