Limitations

Akno is an autonomous-capable memory layer, not a general knowledge-management platform. These are the current capability boundaries that materially affect how it should be used.

One owner or one trusted shared owner

Akno assumes a person and their agent share authority over one knowledge base. It has one write handle and transport-level allow lists, but no user accounts, per-page ACLs, tenancy, approval roles, or attribution model for mutually untrusted collaborators.

This works for a personal memory or a trusted household/team folder with one shared policy. It is not enough for an organization where different people must be unable to read or change each other’s material.

macOS host

The core and CLI target macOS. Document extraction uses PDFKit, Vision, and textutil; background operation uses launchd. The protocol and client packages are portable, so a container or another machine can call a macOS-hosted service, but the full Akno runtime is not a supported Linux or Windows installation.

Conservative entity resolution

Akno has an evidence graph and bounded multi-hop traversal. It resolves identities from exact signals: canonical slugs, explicit aliases, unique titles and basenames, authored links, akno.about, document ownership, and optional contextual choice among already-known same-name candidates.

It does not perform open-ended named-entity recognition or infer arbitrary relationships from prose. Duplicate merging always requires allowed-folder policy. Exact discovery accepts an authored alias or a narrow graph-backed signal: at least two distinct current attributes resolve exactly to a canonical entity with a multi-token name, and the candidate title contains that complete name.

Optional semantic discovery is deliberately narrower than general entity resolution. It compares only complete, compact, synthesize-opted sibling pages in an allowed folder. A permissive embedding prefilter plus a strict classifier passed the invented development corpus and independently reviewed frozen split; the held-out path passed five stable runs with no false positives. That classifier only adds a candidate. The lossless planner, independent verifier, and curator must still agree that the page has no useful separate scope before any write. Repeated templates, scoped subpages, events, procedures, and uncertain pairs stay separate.

This avoids dangerous false joins, but large pages, cross-folder duplicates, loosely written pages outside the qualified classifier contract, and identities that require open-ended reasoning remain outside semantic merge discovery.

Retrieval is not omniscience

recall and answer can only use indexed and permitted evidence. Empty retrieval does not prove that a fact is false or absent from the world. Documents with failed extraction, ignored paths, disabled model roles, ambiguous identity, tight budgets, or unavailable endpoints may reduce the searchable evidence.

Typed empty, degraded, and unavailable states make that boundary visible, but callers still need to handle it. answer verifies support against retrieved evidence; it does not independently verify the evidence against outside reality.

Model quality remains a boundary

Deterministic guards can reject malformed, unsupported, stale, over-budget, or unauthorized model output. They cannot turn a weak inference into an insightful one. Observe, reflect, synthesis, routing, answer generation, and contextual disambiguation remain sensitive to model quality and prompt compatibility.

The qualified OpenAI preset is backed by invented benchmarks, not a guarantee for every private corpus or future provider revision. Run doctor after model changes and use the relevant live benchmark before granting automatic write authority.

The review queue is not universal

Hygiene, synthesis, split, extraction, merge, contradiction changes, broken-link changes, adoption, observation, and reflection use durable plans. Housekeeping identifies broken links, orphan documents, and qualified scalar type drift that are already covered by exact nonterminal plans, but it does not manufacture another item for them. Slug-pattern/depth drift, graph review candidates, report-only conflict findings, and the legacy repair surface do not all become actionable queue items. A human can replace one sealed after-state with plan revise, but there is no snooze decision. An automatic curator can request a bounded correction inside the sealed path/evidence scope, but it cannot expand the plan or turn a report-only diagnostic into an operation.

Consequently, some diagnostics still require a person or future planner to translate a report into an exact operation.

Dependency planning is bounded

A full dream cycle separates all planning from decisions and applies safe items in dependency order. It can replan work invalidated by successful earlier items once. It does not repeatedly loop until every possible transformation settles.

Cross-phase composition, arbitrary same-page proposal merging, and cross-item document-attachment dependency inference remain incomplete. A depth-relocation item does seal and move its own complete attachment set; dependencies between separate items remain bounded. Ambiguous or persistent dependencies are deferred to a later cycle rather than guessed through.

Final run verification hashes the complete indexable tree and treats an unrelated file added, removed, or edited concurrently by a user or sync client as a content-safe unattributed_file_change. It preserves those bytes and fails only the run’s certification. It does not provide an operating-system filesystem snapshot or prevent editors from changing files during planning. A writable full cycle does pin the indexer’s file revision across its initial planner wave and drains queued background reconciliation before decisions. A foreground Akno memory mutation preempts that barrier and makes the dream abort before decisions rather than delaying the user or agent call. Selected single-phase runs remain immediate, and a read-only dry run in another process cannot pause the writable service’s indexer; affected sealed inputs therefore still rely on repeated preflight in those paths. Stronger cross-process or operating-system snapshot isolation is future work, not a requirement for the primary single-service path.

No configurable fail-fast maintenance mode

Autonomous maintenance favors independent progress: one degraded or blocked item does not stop unrelated safe work. There is currently no profile setting that makes the first degraded writable phase abort the rest of the cycle. Installations that require all-or-nothing nightly execution should remain in audit or review and inspect the complete receipt before applying.

Setup does not abstract every provider detail

akno init provides a qualified OpenAI single-endpoint path, a model-free path, and preservation of manually configured specialist roles. The OpenAI preset uses the dedicated Responses adapter, while third-party OpenAI-compatible providers default to Chat Completions and may opt into persistently learned api: auto selection.

Auto-resolution establishes only which generative route accepts every configured model. It cannot prove that a gateway implements strict structured output correctly, that a model honors a requested reasoning effort, or that one provider block can safely mix models requiring different transports. Specialist setups therefore still require understanding role assignment, structured-output support, dimensions, reasoning controls, and degradation behavior; doctor and live benchmarks test those higher-level contracts.

Inference can still look authoritative

Observation and reflection pages are marked derived, kept separate from authored evidence, and ranked lower. They are also append-only, cited, planned, and independently decided. Even so, fluent derived Markdown can look more certain than it is when read outside Akno.

Editors and downstream tools should preserve the role metadata and evidence links. Important decisions should prefer authored or primary-document evidence over a derived observation.

Not an editor, backup, or world model

Akno deliberately does not provide a graphical notes editor, version-control replacement, backup service, general web research agent, or independent source of truth. It can synthesize existing pages using supplied outside facts when the operation and policy authorize that evidence, but it does not autonomously browse the world during ordinary recall or maintenance.

The Markdown files remain the durable product. Keep them readable without Akno and back them up independently.