SillyMemory 0.5.0 (experimental)

SillyMemory 0.5.0 — experimental pre-release

Prepared release boundary: 2026-10-20. Dated notes do not establish publication;
confirm published versions in GitHub Releases.
This boundary captures product changes merged through PR #88 at
cbd27ab470e67af105b2bd56fd3f642c6702cf17, with 0.5.0 installation/host acceptance
merged in PR #90. PR #89's separate BEAM protocol work is not a release prerequisite.
Main promotion, tagging and GitHub Release publication require maintainer authorization.

Supported host: SillyTavern 1.19.0, commit
06bde939fb1e9c4c8d8641d810f0a916b5bce127, with Chromium.
Other host versions and browsers remain unverified.

Changes from 0.4.0

• Query1 searches only the eligible generation anchor in one role-labelled request.
  Recent dialogue stays in generation. Normal input retains at least five eligible
  messages, and a completed reply can retain six to keep an adjacent pair together.
  Continue uses the message being extended; checkpoints still synchronize the full
  completed transcript.
• Bayesian vector/English-text search uses managed OpenAI Small with vector k200,
  native Jev 1.13.0 candidate 200 and final 50, replacing k100/candidate 100. The
  initial and maximum memory allocation remain 4,096 tokens, with the host
  one-quarter-context cap and explicit saved budgets respected.
• Valid chunks are admitted in retrieval order using their final assembled or
  separate size. Selected sources are presented chronologically in one labelled,
  quoted user-context archive. Source identities, original text and recent native
  roles remain intact; archive framing and shared guidance count toward the budget.
• Reranking receives source role, verified host-time context and original text.
  Embedding and English full-text search still use original text only. Unknown
  dates remain unknown; host send times do not establish event dates, and UTC alone
  does not establish a speaker's local date. BEAM batch reference dates are
  evaluation metadata, never product send times.
• Natural guidance supports personalized answers and distinguishes event/role
  matches without a blanket follow-up-question obligation. Native reply formats
  and streaming remain unchanged. Memory preparation failures still stop generation;
  disabling memory restores ordinary native history.
• The shipped SDK is pinned to 0.8.1-dev.38730896042001, the tested candidate-200
  build. No user build step or additional embedding-provider key is needed.

Large embeddings, enriched retrievalText, 8,192-token memory and generated summaries
remain unadopted. There is no extra query-rewrite or generation model phase.
See the product decision,
current defaults and
BEAM evaluation for the separate measured comparisons.
Their historical scores are not a new 0.5.0 answer-quality result or a guarantee.

Installation, update and storage compatibility

Public installs should continue following main through
Extensions → Install extension with https://github.com/lambdadb/SillyMemory
and an empty branch field. Before main promotion, reviewers can select develop
using a disposable host/profile and a fresh test chat, checking the exact installed
commit against the acceptance record. The measured preparation branch was removed
after PR #90 merged; a new branch installation is a new measurement.
Do not install a second copy in an existing profile. The Update button follows
its installed Git branch, not the newest tag or GitHub Release; reload after updating.
Automatic updates remain disabled.

Older experimental collections are not automatically migrated. The role and
conversation-time rerank fields require new keyword indexes. A code update does
not add these indexes to an existing remote collection. Start with a fresh owned
story collection; older collections require explicit owned deletion and
resynchronization before this policy can be used. Preserve original chats and
backups first, and review the affected story branches/checkpoints before any
collection deletion. Do not delete unrelated collections or treat installation
checks as proof of existing-data compatibility. See
the index policy and
storage controls.

Verified browser credentials remain scoped to the native account and exact
endpoint/project. Reload verifies the saved key before restoring the prior enabled
state. Forget key removes it and disables memory. Browser storage is not an isolated
secret vault. No server plugin, SillyTavern credential forwarding or proxy is added.

For rollback, disable memory and preserve native chats first. Pin a known-good
published tag using the release procedure.
Code rollback does not restore remote/local data or reverse collection deletion.

Acceptance and remaining limits

See 0.5.0 acceptance for exact commits, installation/update,
local host/emulator results and retained evidence. These checks are distinct from
historical live LambdaDB acceptance and generated-answer evaluations.

Chat Completion final delivery is checked; other completion formats may report
verification unavailable. Sufficient host context/output space is still required.
Queries above 8 KiB fail explicitly; long ambiguous exchanges can outgrow the recent
window. Correct retrieval does not guarantee correct answers. No arbitrary-provider,
multilingual, group-chat or multi-device concurrent-writer support is established.
A wider candidate pool can increase reranking latency and cost; this preparation
makes no new service-latency, billing or benchmark claim.