Blackboard Shared Memory
Specialists don't talk to each other — they read and write one shared, evolving picture of the problem.
What is Blackboard Shared Memory?
A central knowledge store — the blackboard — holds the current state of the problem. Specialist agents watch it, contribute when their expertise applies, and build on each other's entries without ever messaging directly. The solution assembles itself on the board.
The blackboard pattern is the oldest idea in multi-agent AI, and it has aged beautifully into the LLM era. A structured shared store holds the evolving case: facts, hypotheses, partial results, confidence levels. Specialist agents subscribe to the board and act opportunistically — the document agent posts extracted entities, which wakes the graph agent to link them, which gives the risk agent enough to post a score. No agent knows the whole plan; the board is the plan.
This architecture wins wherever the problem is a case file that accretes evidence over time: a complex claim, a KYC investigation, a clinical workup, a construction coordination issue. Direct agent-to-agent messaging in these settings creates a combinatorial explosion of conversations; a blackboard replaces N-squared chatter with one authoritative state that every specialist — and every human reviewer — reads the same way.
The engineering lives in the board itself: a typed schema for entries, provenance and confidence on every write, conflict resolution when specialists disagree, and a control agent that decides when the board has converged enough to act. Boomlex implements blackboards on vector-plus-relational hybrid stores with full event history — meaning your compliance team can replay exactly how the swarm reached its conclusion, entry by entry.
The coordination loop, step by step
- 01
Post the problem
The initial case — documents, alerts, records — is written to the blackboard as structured, provenance-tagged entries.
- 02
Specialists subscribe
Each agent watches for entry types matching its expertise: entities for the linker, transactions for the risk scorer, images for the vision agent.
- 03
Opportunistic contribution
When an agent's trigger conditions are met, it activates, reads the relevant board region, and posts its findings with a confidence score.
- 04
Knowledge cascades
Each new entry can trigger further specialists — extraction wakes linking, linking wakes scoring — so the solution deepens without central scheduling.
- 05
Conflict arbitration
Contradictory entries are flagged and resolved by an arbiter agent using provenance and confidence, never silently overwritten.
- 06
Convergence & action
A control agent monitors board completeness; when the case crosses its decision threshold, it triggers the downstream action or human review.
Strengths
- Specialists are fully decoupled — add or upgrade one without touching the others
- One authoritative case state; humans and agents read the same picture
- Event-sourced history makes the entire reasoning chain replayable for audit
- Handles asynchronous evidence naturally — new documents just land on the board
- Scales to large specialist teams without N-squared communication overhead
Tradeoffs
- The board is the critical dependency — its schema and infrastructure need real design investment
- Opportunistic activation makes end-to-end latency less predictable than a pipeline
- Requires disciplined conflict resolution or the board accumulates contradictions
- Overkill for simple linear workflows a pipeline would handle cheaper
Every topology has a bill. We tell you what it is before we build.
Reach for blackboard shared memory when…
Where blackboard shared memory earns its keep
Complex-claim case board
Large-loss claims accrete evidence for weeks. Document, imagery, fraud, and coverage agents each post to the claim's board as material arrives, and the adjuster sees one continuously-updated case picture instead of six tool outputs.
ImpactLarge-loss cycle time down 50% with full evidence lineage
KYC / AML investigation board
Sanctions, adverse-media, corporate-registry, and transaction-pattern agents contribute findings to a shared case file, with an arbiter resolving conflicting entity matches before the compliance officer reviews.
ImpactEnhanced due diligence cases closed 4x faster
Diagnostic workup board
Labs, imaging reads, history summarization, and guideline-matching agents post to a patient's workup board as results arrive, assembling a differential the care team reviews in one place.
ImpactComplete workups assembled before the follow-up visit
Site coordination board
RFIs, drone imagery analysis, schedule agents, and safety monitors write to a per-site board, so clash detection and delay risks surface from the combined picture rather than siloed reports.
ImpactCoordination clashes caught days before they hit the field
Grid incident state board
During a grid event, telemetry, weather, asset-history, and crew-status agents maintain one live operational picture, giving operators a single converged view instead of a wall of dashboards.
ImpactRestoration decisions made from one shared truth
What we typically wire together
Ship a blackboard shared memory swarm on your workflow
Tell us the process you want to automate and we'll map blackboard shared memory onto it — orchestration layer, guardrails, and observability included, with timeline and cost estimates.