Company Brain synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From Slack request to scoped action
Company Brain is a self-hostable Slack agent that owns an application-level model/tool loop. It captures company knowledge, supplies retained context to later turns, acts through connected applications and optional sandboxes, and supports scheduled work and operator-reviewed research. This account covers the shipped application at commit ef8a45ef3aed97b32599c7d8007edefab6d14e9a, with an analysis cutoff of 2026-10-14. Its evidence is inspected implementation and attributed design instructions from SRC-1, https://github.com/supermemoryai/company-brain. The whole-system boundary covers shipped responsibility paths, not exhaustive inspection of every leaf. Hosted Supermemory, model providers, Slack, connected services, SDK internals and execution infrastructure remain outside the implementation boundary. No execution trace, passed test result or causal experiment was supplied. Implementation wiring is therefore distinguishable from successful operation and benefit (boundary and evidence, source register).
A signed Slack event resolves a workspace, organization member and conversation surface, then dispatches to the organization's Durable Object. The HTTP acknowledgment precedes the answer. The caller supplies conversation, attachments, identities and deadlines. The agent loads company context, ambient knowledge, learned style, workspace directions and a skill index, assembles actor-scoped tools, and runs the foreground model. The model chooses actions and a terminal proposal; host code controls active tools, step budgets, suspension and finalization. Normal settlement reconciles pending updates before the Slack caller publishes and performs eligible memory writeback. Thread revision checks constrain publication by coordinated Slack turns, not every caller or already-performed remote action (RT-RTE-slack-turn, RT-CLM-turn-fencing).
There is no single model or universal executor. Foreground, triage/fast and entity-research roles use provider models without immutable weight pins; local provider reachability and gateway behavior can alter model selection. Fast calls need not honor the organization's triage override. Inspected client calls establish inference, not parameter training (RT-CMP-brain-main, RT-CMP-triage-model, RT-CMP-research-model).
Connected-app Code Mode exposes discovered methods to JavaScript execution. Native callbacks classify effects: read calls can proceed; classified non-read calls are denied on read-only references or paused for requester approval on writable references. Approval resumes a bounded stored conversation after team, requester, membership, workspace, current-turn and 15-minute expiry checks. This is permission for a particular continuation, not factual verification (RT-RTE-app-code, RT-RTE-approval-resume).
The alternate paths change the guarantee. Direct MCP fallback uses a separate approval predicate: an unknown personal tool name without matching write tokens or destructive annotation can reach execution without that predicate requesting approval. Sandbox commands use regex denials, not complete effect mediation, and do not inherit MCP approval classification. Thus the inspected controls cannot support “every external write requires approval” or a universal background read-only guarantee. Scheduled turns forward the stored payload's read-only flag; research drafts explicitly request read-only operation, but that posture is still scoped to the implemented controls (RT-RTE-direct-mcp, RT-RTE-sandbox-work, RT-CLM-write-approval, RT-RTE-scheduled-task).
From activity to retained context
Interactive capture proposes compact subject documents from conversation and tool observations. Completed eligible callers submit them without per-document human approval. Public-channel observation is a separate producer: it spools messages, asks a fast model for durable clusters, persists those clusters before submission, and retries failed admission without redistilling the batch. Only complete batch submission advances the cursor and deletes local pending material. HUMAN-only provenance, preserved dates and retained rationale are requested model behavior, not independently checked output properties (RT-RTE-knowledge-handoff, MEM-RTE-channel-curation).
Submitted documents are not the same objects as later recalled memories. The application sends text, IDs, containers and flattened metadata to Supermemory, which performs opaque extraction and embedding. queued or done grants submission success, not completed correct extraction. The compatibility adapter does not forward its accepted instant/replacement controls. Later readers consume derived text, liveness fields, buckets and document references; those fields cannot reconstruct the extractor's reasoning or establish every claim's provenance (MEM-OBJ-brain-document, MEM-OBJ-derived-memory, EPI-RTE-memory-admission, EPI-RTE-document-retention).
Recall has distinct automatic and requested operations. Invocation-triggered ambient supply selects accessible person/channel rows, bucketed profiles and a tagged baseline, then shortens each memory to 300 characters. Candidate caps and bucket eligibility make this a partial view, not a complete profile. Requested semantic search ranks and globally caps returned results; tag recall uses lexical metadata matching and estimated token packing; tree navigation follows local document mappings and pages hydrated derived memories. Similarity licenses relevance selection, not truth. Timestamp-only paging and mapping reconciliation have countercases that prevent guaranteed enumeration (MEM-RTE-ambient-recall, MEM-RTE-requested-recall, EPI-RTE-brain-selection).
Local SQL tags, node mappings and generations are access structures, not the remote knowledge store. Nodes reaching 125 live documents can trigger a model partition over at most 160 sampled documents. Remote metadata updates precede local moves; this changes retrieval organization without establishing factual correction. Forgetting is separately requester-gated and limited to conversation write scope plus agent-self. Apply does not locally prove that supplied IDs came from the prior preview. Local reset clears producer/index/research state and advances a generation; it is not complete remote erasure. Entity cache lookup and writes have generation guards, but a web result can still return after reset during inference (MEM-OBJ-memory-index, MEM-RTE-node-splitting, MEM-RTE-memory-withdrawal, RT-RTE-entity-web).
Retaining experience and changing directions
The strongest supported learning-related contribution is implemented automatic trace-fed retention with later consumers. Company documents, public-thread style notes, compacted approval conversations, investigation checkpoints and research history can carry prior activity into future decisions. This is durable behavior-shaping capacity; it does not demonstrate improved capacity for future action. The update consumer and later answering consumer have different authority: retained messages can feed note production, while resulting facts or directions enter later model context (MEM-BAP-retained-context, MEM-OBJ-agent-style, MEM-RTE-investigation-recall).
Style observation skips DMs and channels not positively public. It requests direct HUMAN preference evidence and produces generalized agent-self notes. Later turns receive selected bucketed facets, not the original human evidence or the observer's carried tail. Those notes request voice and operating adjustments below workspace instructions and fixed policy. Workspace prompts, requester-approved skills and model-mediated admin configuration edits have separate admission paths; none is a provider-weight update or demonstrated reliable compliance (MEM-RTE-style-recall, EPI-RTE-style-generation, EPI-RTE-style-admission, RT-RTE-instruction-admission, MEM-RTE-config-edit).
Continuation preserves bounded values rather than recoverable originals. Deterministic compaction preferentially removes old state, skill bodies and older tool output, leaving heads, digests and reload stubs. A skill reload obtains current content, not necessarily the omitted version. Investigation checkpoints retain selected answer/tool material for a matching principal and thread with a six-hour idle TTL. Their verifiedEvidence field can contain answer text selected because any native call succeeded; it is not claim-level verification. Independent research turns isolate full transcripts and return selected products. These mechanisms explain what remains available after context reduction without implying that a digest can recover an evicted result (RT-RTE-context-boundary, MEM-OBJ-thread-investigation, RT-RTE-research-draft).
Research delivery and feedback
The representative multi-step workflow is an operator-triggered research round, not autonomous daily research merely because its source name is “auto-research.” A local lease coordinates planning and concurrent independent draft turns. Planning sees operator steering, roster/context and capped prior-body snippets. Each draft receives a recipient- or target-specific frame and preloaded history, not siblings' newly completed drafts. Reset and lease ownership gate storage. Noncompleted, empty and placeholder products are dropped; excessive length and missing evidence-producing tools are advisory flags, not hard refusal conditions. Stored drafts remain intermediate results (workflow account, RT-RTE-research-draft, EPI-RTE-draft-check).
A reviewer can edit, dismiss or withhold a draft. Explicit send rechecks entitlement, workspace and destination, claims the row, posts the stored body and records successful Slack acknowledgment. Review sources and flags are not posted; the current destination can differ from the review-time label. Failed post results release the claim, but a crash leaving sending is not recovered by the inspected entry point because it rejects that status before reaching the stale-claim helper. Stable UUID posting requests deduplication without proving exactly-once publication (WF-RTE-draft-delivery).
Reception has an implemented consumer. Non-refinement administrative inquiries acquire reactions and bounded reply excerpts for selected sent posts, render them, and supply a fresh model turn whose answer reaches the operator. This is feedback interpretation capacity, not automatic engagement-driven research planning. Refinement is separate and skips reception. A failed thread read paired with a successful reaction read can incorrectly render “no replies” as known engagement; engagement itself does not establish factual correctness (EPI-RTE-outcome-acquisition, EPI-RTE-outcome-frame, EPI-RTE-admin-consumption).
What the architecture establishes
For conversational recall and operator-reviewed drafting, the supported result is a connected acquisition–retention–selection–action architecture with route-specific controls. For guaranteed current facts or evidence-accepted conclusions, the evidence is weaker: hosted supersession is requested but opaque; model claim checking is instructed but unobserved; the operative reply gate checks selected display shapes, not propositions against evidence. An ordinary false sentence can pass, while valid bilingual text can fail. Human publication choice is consequential permission without a required evidence-consuming factual criterion. The product's absolute “not internet” claim also exceeds the implemented source boundary because provider-native entity web search exists (MEM-CLM-service-curation, EPI-CLM-current-memory, EPI-CLM-team-answers, EPI-CLM-evidence-check, EPI-RTE-reply-check, EPI-RTE-reply-selection).
The records do not establish a complete theory-builder assessment. Localized directions are inspectable, but that alone does not establish a formulated explanatory theory (condition 1), decisions depending on its content (condition 2), content-directed criticism with placement of blame and changed reliance (condition 3), or retained criticism shaping another round (condition 4). These are independently unestablished, not evidenced absent. No supplied episode demonstrates that criticism of a consumed theory improved future capacity. Style self-representation and administrative feedback provide relevant partial mechanisms, but a complete reflective-system assessment is not established. Computational control of individual updates likewise does not establish autonomy over every internal theory-building operation (RT-RTE-instruction-admission, RT-RTE-knowledge-handoff, EPI-RTE-admin-consumption).
At the application boundary, the style pathway requests evidence-responsive change toward human-grounded interaction preferences and wires retained directions to later context. Exercised self-improvement remains unestablished: no supplied subsequent operation shows causal dependence on the update, and no comparison demonstrates improvement. Nor do reception interpretation or edited external drafts alone establish self-change. This does not negate wired trace-fed retention or imply that uninspected branches lack criticism (EPI-RTE-style-generation, EPI-RTE-style-admission, MEM-RTE-style-recall, EPI-RTE-admin-consumption).
Concrete changes would alter these assessments: claim/evidence checks consumed by publication or memory admission would strengthen factual acceptance; inspectable hosted extraction and tenant contracts would resolve provider guarantees; a feedback episode tracing criticism through retained revision to later decisions would support iteration; and controlled before/after evidence with independently varied updates would support improvement attribution. Uniform effect mediation, post-inference reset fencing and reachable crash recovery would strengthen their respective operational guarantees. None of those stronger conclusions follows from the current wiring. Report verification found no remaining blocker; profile verification retained one local configuration-maintenance omission, carried below (report verification, profile limits). No additional source reads were made for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation or causal comparison | SRC-1; RT-RTE-slack-turn, MEM-BAP-retained-context, WF-RTE-draft-delivery | Frozen shipped implementation and instructions; no supplied runs | Successful delivery, reliability, activation, measured learning benefit and causal improvement | Registered execution episodes and appropriately controlled outcome comparisons |
| Hosted derivation, initial provenance and tenancy remain opaque | SRC-1; MEM-OBJ-derived-memory, MEM-OBJ-profile-config, MEM-CLM-service-curation | Local client, settings and returned-field treatment; visibility ignored in provisioning | Reliable extraction, semantic deduplication/supersession, embedding provenance and end-to-end organization isolation; uncertainty is not leakage evidence | Frozen provider implementation or response evidence, plus deployment-key and authorization contracts |
| Local effect controls differ across executors | SRC-1; RT-RTE-app-code, RT-RTE-direct-mcp, RT-RTE-sandbox-work | Classified Code Mode, direct predicate and regex shell validation | Universal approval, read-only containment or remote rollback | Complete effect mediation across alternatives and inspected remote/SDK contracts with run evidence |
| Recall and navigation are bounded and may diverge | SRC-1; MEM-RTE-ambient-recall, MEM-RTE-requested-recall, MEM-RTE-node-splitting | Capped/bucketed profiles, ranked and packed reads, timestamp pages and split metadata | Complete profiles, exhaustive enumeration and unconditional tag/tree consistency | Provider metadata normalization evidence, tie-safe paging and mapping reconciliation that distinguishes failed status reads from missing documents |
| Withdrawal does not prove exact-preview application or total erasure | SRC-1; MEM-RTE-memory-withdrawal | Explicit-ID forget, narrower write scope, stale deletion and local reset | Exact prior-preview enforcement, atomic forgetting, remote cascade/re-extraction prevention and whole-memory wipe | Checked preview-to-apply binding, complete store inventory and inspected deletion/reset callers and service semantics |
| Entity return freshness is not fenced after web inference | SRC-1; RT-RTE-entity-web | Guarded cache lookup/write, unfenced web return | End-to-end stale-read prevention across reset | Post-inference generation guard before delivery and evidence covering the reset race |
| Continuation loses originals and may overlabel evidence | SRC-1; RT-RTE-context-boundary, MEM-OBJ-thread-investigation, MEM-RTE-investigation-recall | By-value compaction/checkpoints; full checkpoint rendering not completely inspected | Recoverable omitted tool results, verified answer provenance and complete downstream consumption | Inspected rendering/consumer trace, source-linked claim verification and an actual archive/reload path for originals |
| Publication checks do not establish factual acceptance | SRC-1; EPI-RTE-reply-check, EPI-RTE-reply-selection, EPI-RTE-draft-check | Shape heuristics, advisory tool/length flags and human send permission | Evidence-supported truth, source fidelity and compliance with model self-check doctrine | Named factual criteria and inspectable evidence-consuming decisions controlling reliance or publication |
| Research storage and delivery are not a complete recovery protocol | SRC-1; RT-RTE-research-draft, WF-RTE-draft-delivery | Capped preloaded history, independent siblings, explicit send and caller/helper status mismatch | Exhaustive nonrepetition, coordinated sibling novelty, crash recovery and exactly-once Slack publication | Complete history/selection evidence, explicit sibling coordination, reachable stale-claim recovery and remote deduplication evidence |
| Reception can mistake missing reply evidence for no replies | SRC-1; EPI-RTE-outcome-acquisition, EPI-RTE-outcome-frame, EPI-RTE-admin-consumption | Bounded Slack acquisition and operator-only interpretation; refinement skips outcomes | Reliable engagement absence, automatic reception-driven planner adaptation and exercised criticism/improvement | Failure-aware completeness flags and source identities; a traced feedback-to-update-to-later-operation episode |
| Remaining available leaves and provider models are not fully characterized | SRC-1; RT-CMP-brain-main, MEM-RTE-human-readers, MEM-RTE-debug-capture, RT-RTE-sandbox-work | Named model aliases, bounded HTTP/UI, membership, evaluator and optional sandbox coverage | Immutable model equivalence, exhaustive authorization parity, complete durable sandbox inventory or system-wide absence of acceptance/criticism | Frozen model identities and focused inspection of the relevant omitted consumers, admission gates and provider contracts |
| Extraction-configuration replacement is omitted from the partial curation inventory | SRC-1; MEM-RTE-profile-provision, MEM-OBJ-profile-config; profile verification limit | Accepted records establish subsequent replacement of retained guidance/buckets; profile inventory leaves refresh unresolved partly because it is not factual text | Complete mechanism-level configuration-maintenance comparisons; the independently supported evolve union remains valid | A revised comparison inventory classifying recorded configuration replacement as evolution without requiring factual-text revision |