Institutional memory
Knowledge that keeps its source when the person leaves.
Your company knows a lot. The problem is where it lives: in a few experienced people, in thousands of documents nobody indexes, in chat threads that scroll away. We build systems that capture that knowledge as facts with a source, a date and an owner, so people get answers they can check instead of answers they have to trust.
In plain terms
An institutional memory system, or AI knowledge management system, captures what an organization knows as individual facts that carry their source, their date and who owns them, so people and software can find and trust an answer without hunting through documents or asking the one person who remembers. Blocpod builds these systems with permissions preserved and contradictions sent to a person to resolve.
The company knows. Nobody can find it.
Every organization has a memory, and it lives in the wrong place: inside a few experienced people, across thousands of documents nobody indexes, and in chat threads that scroll away. So decisions wait while someone reconstructs history. New people take months to become useful. And when the person who understands a process is unavailable, the process stops.
Search does not solve this, because search returns documents and you still have to read them. Generic AI assistants do not solve it either, because they answer from a blend of your files and their training with no way to tell which, and no idea whether a document is current, superseded or restricted.
Blocpod memory systems are built on claims, not chunks. Facts are extracted with their source, their date and their owner. Contradictions are quarantined for a human to resolve instead of averaged into a confident wrong answer. Permissions travel with the knowledge. What you get is context you can act on and defend.
Good candidates
If your team does any of these, it is probably worth a look.
Context reconstruction before decisions
Escalations, renewals, disputes and reviews that start with someone digging through five systems to find out what happened.
Policy and precedent retrieval
What is our position on this, and when did we last decide it? Answered with the policy, the precedent and the date.
Operational how-it-works knowledge
The processes that only run because one person remembers them, captured as verified steps with owners.
Decision records
What was decided, by whom, on what evidence, so nobody re-litigates it from memory.
Onboarding and handover
New people and successors working from the organization’s memory rather than the previous holder’s inbox.
Evidence-backed recommendations
Briefs and recommendations for operators where every line carries a citation to its source.
What Blocpod builds
The whole path from request to result, not a chatbot bolted on.
Connects to
- Google Workspace and Microsoft 365
- Slack and Teams
- Notion, Confluence and wikis
- Shared drives and document stores
- CRM and ticketing history
- Meeting transcripts
- Agent runtimes such as Claude Code and Codex via compiled context
- IngestSources connected with permissions intact.
- ExtractClaims with source, date, owner and evidence span.
- ReconcileDuplicates merged; contradictions quarantined for review.
- BoundaryConflicts and sensitive knowledge routed to owners.Human decides
- ServeDigests and context packs for people and systems.
- RecordEvery answer cites; every change is logged.
- 01
Source ingestion
Documents, threads, transcripts and records ingested from named systems, with permissions and ownership preserved at the source level.
- 02
Claim extraction with provenance
Atomic facts extracted with their source, date and evidence span. Agent-generated text is never treated as a fact.
- 03
Contradiction handling
When sources disagree, the conflict is quarantined and surfaced to a person. The system does not pick a winner silently.
- 04
Living digests and context packs
Compact, current summaries per account, project or process, and task-specific packs compiled with hard budgets for people and agents.
- 05
Permission-aware access
Who can see what is enforced by the system, not by hoping the model forgets. Inference, synthesis and source are distinguished.
- 06
Freshness and ownership
Every claim knows how old it is and who owns it. Stale knowledge is hedged or flagged, not presented as current.
Who does what
What the AI does. What stays with your people.
The system / does the work it is allowed to do
- Extracts claims from sources with provenance and dates
- Detects duplicates and contradictions
- Renders living digests and compiles task-specific context packs
- Answers questions with citations and freshness
- Compiles context for engineering and operational agents
- Flags stale or orphaned knowledge for owners
Your people / make the decisions that matter
- Own knowledge domains and resolve contradictions
- Decide what is authoritative when sources disagree
- Set permissions and sensitivity classes
- Approve changes to the record of decisions
- Judge when a recommendation becomes an action
Common failure modes
Why these projects usually fail elsewhere, and how we avoid it.
Most of these failures happen before any code is written. We rule them out in the diagnostic. If you have lived through one already, tell us; it shortens the conversation.
Chunks instead of claims
Retrieving text fragments and letting a model summarize them produces fluent answers with no provenance and no notion of freshness.
Contradictions averaged away
Two documents disagree and the model blends them into something confident and wrong. The right behavior is to stop and ask.
Permissions as an afterthought
A knowledge system that leaks a restricted document to the wrong person is a breach, not a feature.
Agent output stored as fact
Once a system’s own summaries are treated as sources, errors compound. Inference must stay labeled as inference.
Boiling the ocean
Ingest everything and you get noise with citations. Start with the domain behind one expensive workflow.
Implementation and measurement
How it works: Diagnostic, Pilot, Production, Expansion.
Every engagement starts with one workflow and a target we agree on before we build anything. The pilot runs on your real workflow with real permissions. Production is measured against where you started. We only expand when the numbers say so.
What we measure
- Time to reconstruct context for a decision or escalation
- Answer precision on audited questions, with citations checked
- Contradictions surfaced and resolved
- Onboarding time to first independent work
- Dependence on named individuals for a process
- Stale-knowledge rate in served answers
Related systems
Systems we have built the same way.
Blocpod systems shown here are internal or public products. Client case studies are added as disclosure is authorized.
- Blocpod system / v0IssuerOS: governed investor communications where 17 agents recommend and humans decideA multi-tenant operating system for creating, verifying, approving and launching investor-facing communications. Source to claim to content to approval to release, with a hash-chained audit log per organization.
- Blocpod system / Pilot-stage internal systemFounderOS on Meridian: a governed executive office running on verified contextMissions, specialist agents, compiled context packs and approval boundaries for a founder's real workload. External actions are draft-first, memory is gated, and every fact carries its source.
Questions buyers ask
Straight answers to the questions people actually ask.
Is this a RAG chatbot?
It uses retrieval, but it is built on extracted claims with provenance, dates and owners rather than on text chunks. Contradictions are quarantined, permissions are enforced, and answers cite. Meridian, our own context operating system, is the reference architecture.
Does our data train models?
Provider terms and any model training questions are settled explicitly in the engagement and written into the system design. We do not make a blanket public claim here; we make a specific commitment in your scope.
How are permissions handled?
Permissions are preserved from the source systems and enforced by the memory system’s server, never by the model. Sensitive knowledge classes route to owners rather than being served broadly.
Can other systems use the memory?
Yes. Compiled context packs can brief operational agents and engineering runtimes so that automation works from the same verified knowledge as your people.
Where do we start?
With the knowledge behind one expensive workflow: the escalation that takes an hour to research, the process that stops when one person is out. That is what the diagnostic scopes.