Changelog

Published 2026-08-19 · Last reviewed 2026-09-11 · Applies to SPM public releases

Patch releases ship reliability fixes continuously; minor versions accompany measured capability changes published on the Benchmarks page.

2026-09-11

• Lane ordering is live. explicit_first now orders the final recall evidence set: deliberate saves lead, passive proxy captures follow — across both candidate evidence and source spans. observed_only answers only from proxy-captured content. This completes the lane rollout announced on 2026-09-07.
• Recall responses now include an evidence_decision: a machine-readable verdict (supported, insufficient, contradicted, or unknown) with the evidence and source IDs it is based on, an authority label, and — when evidence is declined — why it failed support (for example a CJK coverage floor). Treat supported as settled; treat insufficient as "keep looking", not as an answer.
• Memories become recallable sooner and stop getting stuck. New database indexes cut the wait between saving a memory and it being ready to recall, and empty extraction payloads are now skipped instead of parking in a permanent retry state.
• Console dashboard improvements: usage views paginate server-side and refresh automatically instead of loading one giant slow query.

2026-09-07

• Faster, steadier answers through the Provider Proxy: memory lookups on the proxy path now use the bounded fast lookup by default, memory captures run in parallel with a hard deadline instead of one-by-one, and unchanged history is no longer re-saved on every turn. Every proxied request logs its per-stage memory timings, so slow memory behavior is visible instead of silent.
• Chained Responses requests are no longer a memory blind spot: calls that continue a provider-side thread (previous_response_id) now save their new input and can receive recalled context, while compression stays off for provider-managed threads.
• Token estimates are fairer for CJK text: Chinese, Japanese, and Korean content is now estimated per character instead of by byte count, so budgets and billing estimates stop undercounting CJK conversations.
• Anthropic tool pairing is protected when memory is injected: recalled context never lands between a tool call and its result.
• New lane_policy parameter on the MCP recall tool and recall APIs: explicit_first (default) ranks deliberate saves ahead of passive proxy captures, blended keeps one shared pool, and observed_only answers "what did I say in chat" questions from proxy captures. The parameter is accepted today; the server-side lane ordering itself activates after a paired evaluation gate.

2026-09-05

• End-user partitions: one API key can now serve many end users with isolated per-user memory spaces. Give the key the memory:partition scope and send a stable opaque user_partition on remember, recall, and read calls. Partitions are fenced across writes, recall, evidence reads, and read tokens.
• Stored Memories now shows a short topic label derived from each source's own content (plus its type: Input, Tool output, or Assistant) instead of a bare source ID, so you can tell memories apart at a glance.

2026-09-03

• Deep recall now always weighs your saved facts alongside raw conversation windows. Previously, a large collection of windows could occupy the whole review budget, so answers to multi-part questions could miss facts that were actually saved. Saved facts are now guaranteed a share of every deep review.
• Memory evolution is more reliable: long-running background updates no longer stop mid-write, and a batch that cannot be processed is now skipped with an audited record instead of stalling the queue.
• Fixed a server error that could intermittently reject memory saves and deletions.

2026-09-01

• auto recall now returns a deeper result only after Deep recall completes with sufficient evidence. Partial or interrupted gathering stays a clear no-evidence result instead of becoming an answer.
• Deep recall keeps gathering when the selected evidence is still missing a requested premise, even if the words in the question already appear in the current evidence. This improves multi-part questions without weakening refusals for absent or look-alike facts.

2026-08-29

• Deep recall now uses a search index prepared as memories are saved. Large memory collections no longer need to rebuild that index during each recall, so repeated questions are more predictable.
• Readiness checks are stricter: Deep recall only uses a memory collection when its saved sources are fully searchable. If processing is still underway, SPM returns a temporary, retryable unavailable result instead of silently using partial evidence.
• Interrupted memory deletion is safer to retry. Repeating the same deletion continues its cleanup and returns the original receipt ID after completion, rather than treating unfinished work as complete.

2026-08-26

• Provider settings rebuild: the model catalog now auto-admits providers that models.dev lists with a working HTTPS endpoint and a documented credential variable — no waiting for a review cycle — while model pricing, context windows, and capability markers refresh automatically with the catalog. Custom providers can now accept a manually typed model ID when the upstream has no model-list endpoint.

• Published @spmos/local-proxy@0.1.2: old tool outputs are now replaced by short retrievable stubs once safely stored in memory — on all three supported protocols, including Anthropic Messages. Tool pairing is never broken and prompt-cache breakpoints are never touched.

• The hosted Provider Proxy now applies the same tool-output stubbing to Anthropic Messages traffic, matching the OpenAI-compatible paths.
• The console is reorganized around what you want to do (Overview, Memory, Usage, Developer, Account), and Settings now lets you tune memory behavior yourself: default recall depth, capture switches, ambient retention period, compression defaults, and token budgets.
• Recall is better at real questions in Chinese and English, and no longer surfaces a saved copy of your own question as if it were the answer.

2026-08-25

• The console is reorganized around tasks: Memory (Stored Memories, Recall Playground, Activity), Usage, Developer (API Keys), and Account.
• Recall gains a History tab: past recalls filter into answered and declined, replacing the separate Refusals page (old links redirect).
• Settings is now tabbed: General (username and display name), Providers, Memory, Proxy, Security (password), and a red-boxed Danger zone (clear memory, close account).
• New account preferences in Settings: default recall depth (auto/fast/deep), automatic capture toggles for your messages and assistant replies, proxy compression default, and input/recalled token budgets. Per-request API headers still override them for a single request.

• Fixed the Recall Playground error display: a client-side schema that lagged behind the server could dump a raw validation error for newer refusal reasons. Refusals now always render as a clean card with a plain-language reason, and unexpected responses show one short line.

• Tool-output elision savings now appear in request receipts: agentic requests show the elided_tool_output continuity state with the original and forwarded token counts, so removed old tool outputs are visible as real savings instead of a passthrough zero.

• Deep recall now keeps supporting premises in the evidence list even when the answer text is intentionally concise. Agents can use the attached read tokens to retrieve the exact original wording for each premise. Refusals for look-alike evidence are unchanged.
• Deep recall diagnostics now explain how much work was used, whether evidence coverage completed, and the provider token and latency cost. Slow selection times out sooner so recalls fail clearly instead of hanging.

2026-08-24

• Published @spmos/local-proxy@0.1.1 to npm: one-command install now ships the continuity contract (x-spm-continuity-state, two most recent exchanges protected, removal only on proven recall) and the stricter capture hygiene (reasoning and tool-argument blocks never become memory).
• auto is now the default recall depth: SPM answers with the bounded fast lookup first and only escalates to deeper multi-round gathering when that lookup is not confident enough. A question that clearly names something not in memory (for example a marker you never saved) is refused without paying for deep gathering.
• Deep recall is cheaper and faster: it stops as soon as the gathered evidence covers every part of the question, sizes the evidence sent for selection to the question's complexity, and only sends newly gathered evidence in later rounds instead of resending the full catalog.
• Memory now consolidates related facts into evolving observations. When something changes — a preference, a configuration, a decision — SPM records the current state with its provenance and keeps the earlier state as history instead of silently overwriting it. Each observation accumulates a proof count as more saved facts support it.
• Re-saving a large changed document is faster: only the parts that actually changed are processed again.
• Fixed recall of pasted code and exact identifiers: asking about an exact function, path, or token name now reliably returns what you pasted, in both auto and deep modes.
• Request receipts now show the requested and effective compression mode and the continuity state, so a request that compressed nothing is labeled truthfully instead of showing a misleading zero.

2026-08-23

• Added explicit fast and deep recall modes: depth=deep performs multi-round evidence gathering for harder questions and reports rounds, provider usage, and latency; fast remains the bounded default.
• Tightened abstention: recall now requires evidence with semantic support for the question beyond shared vocabulary, so unanswerable questions return an explicit no-evidence outcome instead of a look-alike answer.
• The console Dashboard now shows estimated input-token savings (clearly labeled reference estimate, not billing) and a one-click shareable savings receipt.
• Fixed Google and GitHub sign-in completion, account-closure requests, and the post-closure redirect so closed accounts land on sign-in with a clear notice.

2026-08-21

• Added Local Proxy for users who want upstream provider credentials to remain on their own machine.
• Improved continuity in long conversations and made recalled evidence more focused.
• Improved provider compatibility, provider and model setup, and usage reporting.
• Updated the public product identity to SPM — StellarPath Memory Operating System.

2026-08-20

• Added profile management, username sign-in, and Google and GitHub sign-in.
• Added multi-model provider management and model discovery for Custom BYO providers.

2026-08-19

• Opened registration for users with a confirmed email address.
• Added plan usage controls and memory and account deletion workflows.

2026-08-14

• Improved memory deletion reliability and retrieval quality.

2026-09-05

• Added end-user memory partitions: one application key can now serve multiple users while keeping each user's saved memory separate. The dashboard API-key flow exposes the memory:partition scope, and MCP requests can carry a stable user_partition.