How it works · MCP & SDK

A memory network is a group
of people and agents that learn together.

What one agent learns doing the work, every agent in the network can use, with the trust it has earned. Governed by default through trust and validation rules, with human review where you require it. Nothing you write leaves your network.

01System overview

One network. Every agent.

Different tools, different people, one network. Agents read by search, contribute by create and enrich, and signal by feedback. The network ranks, validates and reconciles; agents never coordinate with each other directly.

behind your firewall

Claude Code

user · alice

Cursor

user · ben

Windsurf

user · carla

+ n more

agents · ci · scripts

mcp · sdk

memco instance

continuous · autonomous

Hybrid retrievalvector + bm25 + trust
Trust modellingbayesian · decays
Validation rulesagentic or human review, by config
Reconciliationconflicts resolved · evidence kept
read · write

Your network

private · default

your team’s knowledge · never leaves the tenant

tools · search · create_memory · enrich_memory · share_feedbackauth · sso · tenant isolation

fig. 01 — agents on the left, one memco instance, your network on the right

Four agent nodes inside your firewall, Claude Code, Cursor, Windsurf and more agents, CI and scripts, connect over MCP or the SDK to a single memco instance running hybrid retrieval, trust modelling, validation rules and reconciliation. The instance reads and writes your network, which never leaves the tenant.

02Knowledge domains

Two kinds of work. One connection.

A domain is what a network learns about: what an insight can be about, what its tags mean, and which agents contribute. We ship two, coding and knowledge work. They are fully isolated: a search in one never returns the other, and both are available at the same endpoint.

● codingagents · claude code · codex · cursor

Software development.

API patterns and conventions, error-handling strategies, architectural decisions, dependency-specific behaviours, testing patterns, deployment procedures. Tagged by language, framework and repository.

the domain the coding plugins write to

● knowledgeassistants · claude · chatgpt · cowork · grok

Knowledge work.

Process knowledge, policy and compliance, stakeholder and organisational context, domain terminology, analytical frameworks, procedures. The judgement that would otherwise disappear into a chat.

enterprise customers can define their own domains with us

03Agent interface

Core calls.

Agents talk to memco over MCP, or through the SDK when you own the agent loop. The core read, contribute, feedback cycle is four calls.

01 · readat task start, on errors

search(query, context)

Retrieve relevant insights

Searches your network. Ranked by a hybrid of vector similarity, full-text match and trust-weighted evidence, not similarity alone.

hybrid retrieval · vector + bm25 + trust

02 · writeon discovery

create_memory(content)

Store a new insight

Triggered organically when an agent discovers an undocumented behaviour, a workaround, a convention. No prompt required. Enters the store with an estimated trust score.

organic · a side-effect of work

03 · writeon outcome

enrich_memory(id, content)

Refine an existing insight

Adds context to an insight the agent has already read: for example, attaching the outcome of a task to the rule that informed it. The primary path for in-session refinement.

refinement · keeps insights current in place

04 · signalafter retrieval

share_feedback(id, score)

Was the insight useful?

Positive and negative scores both feed the Bayesian trust model. The closing edge of the loop: feedback from one session shapes ranking for every session after it.

closes the loop · evidence updates trust

04Trust modelling

Ranked by evidence, not similarity alone.

Every insight carries a trust score backed by a Bayesian distribution, not a scalar. Each retrieval, each relevance signal, each enrichment updates that distribution.

01

Bayesian, not scalar

Trust is a probability distribution. Each piece of evidence updates the posterior; confidence reflects both the level and the weight of accumulated signal.

02

Decays without signal

Insights that stop being retrieved or stop earning positive feedback gradually lose confidence. Nothing sits on the shelf forever because it once worked.

03

Exploration, not just exploitation

Retrieval samples from the distribution. High-confidence insights surface reliably; lower-evidence insights are selectively included so they can earn signal.

04

Server-side only

Trust is operational. The agent never sees scores and the user never sees them in chat.

trust distribution · per insight

 trust →      0        .25       .50       .75        1
 ─────────────┼─────────┼─────────┼─────────┼─────────┼

 new          ▁▂▃▃▃▃▂▁
              wide · low weight

 emerging               ▂▄▆▆▄▂
                        narrowing

 validated                            ▂▆█▆▂
                                      tall · heavy
An insight's confidence is a distribution that narrows with evidence.Diagram of three trust distributions on a 0 to 1 scale: a new insight is wide, an emerging insight narrower, a validated insight a tall narrow peak near 0.9.

05Governance

Governed by default. Human review where you require it.

Trust ranks what has proven useful. Three more mechanisms decide what may be served at all, what must never drift, and what happens when insights disagree.

● validation rulesper network · set by admins

What needs a human look

A network's admins decide what is served straight away and what waits for review, from nothing to everything. Rules can be judged by an LLM against the admin's criteria, so a rule like “anything touching payments” needs no list of keywords.

moderation queue · only where a rule says so

● policiesmaintained by people

What must not drift

Knowledge the network treats as always true and never scores. Policies are written and maintained by people, served to every agent, and never decay. They are for the rules that must hold whatever the evidence says.

always served · never re-ranked

● reconciliationcontinuous

When insights disagree

When two insights contradict each other the network resolves the conflict where the evidence allows, flags it where it does not, and keeps the trail either way. Nothing is silently overwritten.

resolved or flagged · evidence kept

06Knowledge lifecycle

The loop, end to end.

From the moment an agent contributes, the network takes over. Governed by default through trust and validation rules, with human review where you require it. Operators run continuously and write back into the same store the next search reads from.

queue · prior · index

Ingest

create or enrich lands, scored and indexed

rules · review where you require it

Validate

served, or held for a human look

vector · bm25 · trust · sampling

Retrieve

low-evidence insights get their chance

synthesise · reconcile · prune

Curate

operators run continuously

posterior · re-rank

Update

every share_feedback re-scores

fig. 02 — the loop never breaks

07Networks

Private by default. As many as you need.

Networks are logically isolated in storage and ranked separately at retrieval time. Nothing written in one reaches another unless the networks are arranged so that it does.

● your networkisolation · per tenant

Organisation memory

Everything your agents contribute lands here, and it never leaves your tenant. Other organisations cannot read it, and memco does not read it. Not shared with other customers, not aggregated, not used for model training.

every memory call operates exclusively against it

● hierarchyparent · child

Networks in a hierarchy

A child network reads from its parent; writes stay in the child. That is how a team seeds another team, how an integrator seeds its customers, and how roles or product lines get their own memory without losing what is common.

read flows down · writes stay local

● always onboundary · a firewall you control

Nothing leaves

There is no pool your insights drain into. What your agents learn stays inside the network that wrote it, and is never aggregated with anyone else’s. Memory is owned by you.

no setting to turn on

08Portability

Your agents learn. Your models stay the same.

Nothing changes in the weights. The learning lives in the network, so any agent on any model uses it, and it moves with you when you change model, tool or vendor. We have measured learning carried from one model to another.

arXiv preprint

learning carried between two modelsτ-bench bankingarXiv:2607.22157

Integration evidence — the stack memco runs against today

Models
ClaudeGPTGeminiLlamaDeepSeekQwen
Agents
Claude CodeCodexCopilotClaude DesktopChatGPTGrok
IDEs
CursorWindsurfZedJetBrainsVS Code

09Deployment

Two ways to run it. Same interface.

Both keep your organisation’s memory isolated and expose the same MCP and SDK surface.

option a · fully managed

Managed, on memco’s cloud.

Runs on memco’s managed service. Your organisation’s memory is logically isolated: every read and write is scoped to your networks, and role-based access controls who can see and change what. Agents connect over HTTPS from behind your firewall.

hostingmemco cloud
org memorylogically isolated
accessrole-based
time to deployminutes
option b · self-hosted

On-premises, in your infra.

A private memco instance runs entirely behind your firewall. All organisation memory stays on customer-controlled infrastructure. Nothing crosses the firewall.

hostingcustomer-controlled
org memorystays on-prem
interfaceidentical · mcp & sdk

10Setup path

Running in one session.

Install the plugin in your coding agent, sign in, and the agent has memory. The quick start covers Claude Code, Codex, Cursor and Claude Desktop, and a two-agent tutorial that shows the first contributed and reused insight.

claude code — pluginmarketplace · memcoai/marketplace
/plugin marketplace add memcoai/marketplace
/plugin install shared-memory@memco

11Personal memory

Your preferences. Every agent.

Your preferences and strictly personal knowledge, kept out of the shared network and available to every agent you use. Underneath it is a network with its own domain and access rules; to you it is the feature that keeps the team's memory clean.

plugin · personal-memory@memco · or mcp · https://spark.memco.ai/mcp-personal

Personal memory on the team page

12Product FAQ

The technical questions, answered.

See it on your own repos.

Start a teamTry memco free

the loop

benchmarks · product · research

A short dispatch on shared memory for AI agents — the numbers behind the product, what we're shipping, and the research we're reading. No filler.

Unsubscribe anytime · no spam