Professor Synapse v3.3.0

Plugin release. Aligns the git tag with the plugin's own version (plugin.json = 3.3.0) so the Synaptic Labs marketplace pins to a tag that matches what installs — consistent with the other plugins (tag == plugin version).

Same code as the prior v2.5.0 repo release (no-query recall fallback + WSL/UTF-8 fixes); this tag simply gives the plugin its own versioned release line. Future plugin releases continue from 3.x.

Install

/plugin marketplace add ProfSynapse/synaptic-labs-plugins
/plugin install professor-synapse@synaptic-labs

🤖 Generated with Claude Code

v2.5.0 — No-query recall fallback + WSL/UTF-8 fixes

No-query recall fallback + WSL/UTF-8 fixes

A bare summon used to come back empty. recall/brief with no --query returned nothing even when the agent had plenty stored — the --agent flag scopes a query, it doesn't dump records. Now a bare summon hands back the agent's most relevant recent memory instead.

What changed

• No-query recall fallback (_recent_hits). With no query terms, recall surfaces the agent's recent long-term records, reranked by recency fused with the same kind/confidence weights as keyword search — so a high-value fact/lesson can edge out a slightly newer note, but only recent records are ever in play (RECENT_DEFAULT = 8).
  • brief (the summon prefetch — summon.py auto-fills the agent's triggers as the query) falls back whenever nothing matched, returning the set under a new recent key. This fixes the common case: a fresh agent whose triggers don't lexically match its own stored memory.
  • recall (explicit search) falls back only when no query was supplied at all. An explicit query that simply misses still returns an honest empty.
  • Fallback hits carry "why": "recent (no query match)" and are not graph-wired (no query, no Hebbian event).
• Backup write fix: shutil.copy2 → copyfile — copy2 copies metadata via os.utime/chmod and raises PermissionError on Google Drive / WSL mounts. (Skill + Claude flavors; the plugin already had it.)
• UTF-8 stdio: emoji in memory output no longer crash with a cp1252 UnicodeEncodeError on Windows.

Versions

• Portable skill & Claude flavor: v2.5.0 — 62/62 tests pass in each.
• Claude Code plugin: v3.3.0 (installs via marketplace from main; same fallback feature).

Install

• Claude Desktop skill: download professor-synapse-claude.zip → Settings → Capabilities → Skills → Add new skill.
• Portable skill: download professor-synapse-skill.zip and drop the folder into your assistant's skills dir (.claude/skills/, .codex/skills/, …).
• Claude Code plugin: /plugin update professor-synapse@professor-synapse (or install from the marketplace).

v2.4.0 — Two flavors + portable in-place self-update

Two skill flavors

The skill now ships in two self-contained flavors:

• Claude flavor (professor-synapse-claude/, distributed as professor-synapse-claude.zip) — keeps the Claude-app machinery: mandatory packaging workflow, file-operations.md, rebuild-protocol.md, /mnt/skills paths. Updates prepare a merged package you install via "Copy to your skills."
• Portable flavor (professor-synapse-skill/) — drop the folder into any skills directory (.claude/skills/, .codex/skills/, …). Edited in place.

Portable self-update (new)

The portable flavor updates itself in place. Ask Professor Synapse to "check for updates" and it fetches the latest release, merges it, and writes the new files straight over its own install — preserving your memory/ store, custom agents, and Learned Patterns. You don't download or move anything.

Fixes

• update.sh no longer breaks after the folder rename (flavor-aware source path).
• .gitattributes (eol=lf) stops CRLF churn on Windows/Drive mounts.

Both flavors: 62 memory + 13 summon tests green.

---

Claude install: download professor-synapse-claude.zip below → Claude → Settings → Capabilities → Skills → Add new skill → upload the zip.
Portable install: copy the professor-synapse-skill/ folder from the repo into your assistant's skills directory.

v2.3.0 — Actionable verify loop

Closes the confidence/verify loop introduced in v2.2.0. A low-confidence guess no longer stays a guess forever.

scan → unverified

scan now returns a bucket of below-high records that carry a verify path (a known way to firm them up), sorted by reliance — the shaky things the work keeps leaning on surface first.

reconfirm verb

reconfirm --id <id> [--confidence high|medium|low] [--source "what confirmed it"] [--verify ""] [--replace-source]

Folds new evidence into source (appended to the existing trail by default), adjusts confidence up or down (disconfirming evidence is valid), clears or rewrites the verify path, resets the staleness clock, and logs a reconfirm low→high (…) event.

The loop: scan shows what's worth confirming → you check it → reconfirm promotes (or demotes) it with the evidence attached, and it drops off the unverified list.

Tests: 62 memory + 13 summon, stdlib-only.

Install: download professor-synapse.zip and upload it in Claude → Settings → Capabilities → Skills. Existing installs: run scripts/update.sh (preserves your memory/ store).

v2.2.0 — Save-decision policy, confidence basis, self-check

Builds on v2.1.0's programmatic summoning with a smarter, more self-aware memory layer.

Confidence carries its basis

A record can now record why it's at its confidence level: source (evidence held), verify (the upgrade path — what evidence is available and how to get it), and unknowns (genuine gaps). All three are full-text indexed and surfaced in recall, so the agent justifies a level and offers to confirm a shaky one instead of asserting it. A low-confidence inference now ships with its own path to becoming high-confidence.

"When to ask vs. just save" + a real signal

A save-decision policy (told-vs-inferred / destroys-contradicts-exposes → just-save / save-low-confidence-and-narrate / ask-first), backed by a write-time duplicate/contradiction check: the new read-only check verb and a non-blocking ⚠ advisory on record flag near-duplicates and high-confidence conflicts so the agent consolidates or confirms.

Committed self-check

references/self-check.md — a repeatable PASS/FAIL verification of an install (version, summoning, the memory loop, both test suites), with a paste-in prompt for Claude Desktop.

Programmatic agent summoning (baseline from v2.1.0)

scripts/summon.py assembles a boot package — full persona + recalled memory + loadable resources — in one call.

Tests: 55 memory + 13 summon, stdlib-only.

Install: download professor-synapse.zip and upload it in Claude → Settings → Capabilities → Skills. Existing installs: run scripts/update.sh (preserves your memory/ store).

v2.1.0 — Programmatic agent summoning

Adds programmatic agent summoning plus the memory-system advances accumulated since v2.0.0.

Programmatic summoning (scripts/summon.py)

One command resolves an agent (exact slug or fuzzy match) and returns a boot package whose stdout IS the summon: full persona + memory recalled inline (reinforcing by default, stamped to the agent) + auto-extracted loadable resources. --query falls back to the agent's triggers; --json for structured output. summon-agent-protocol.md and the SKILL.md summoning section are now script-first.

Also in 2.1.0 (since v2.0.0)

• Knowledge graph — co-recall affinity + spreading-activation recall; recall/brief --query reinforce by default; weights decay with a half-life.
• "Use it or lose it" cleanup — last_used resets on reactivation; scan flags stale_longterm; new forget verb; well-wired records spared.
• Richer memory body + lesson kind — goal/outcome/constraints/confidence; full-text indexed; multiplicative ranking nudges; in-place schema migration.
• Reasoning-first memory guidance — protocol teaches how to interpret recall (why, constraints, confidence) and pick a record kind, not just commands.

Tests: 44 memory + 13 summon, stdlib-only.

Install: download professor-synapse.zip and upload it in Claude → Settings → Capabilities → Skills. Existing installs: run scripts/update.sh (preserves your memory/ store).

v2.0.0 — Memory baseline + script-driven updates

Major release marking the agent-tagged memory architecture and the script-driven update pipeline as the new baseline for Professor Synapse.

Highlights

• One-command updates — scripts/update.sh detects the latest release tag, downloads the canonical codeload tarball, preserves your memory/ store and custom agents, flags only the files needing a hand-merge, and rebuilds the agent index. --check, --ref, --out, --force.
• Legacy HTML scrapers removed — fetch-github-file.sh / github_blob_parser.py and the html2text dependency are gone. The codeload tarball is the only update fetch path.
• Memory system (baseline) — shared, agent-tagged store (scripts/memory.py): working memory + SQLite long-term + change log, ranked-fusion recall (FTS5 + BM25 + recency/kind via RRF, with a LIKE fallback), the one-shot brief prefetch, the 🧠 Memory Keeper agent, and a 25-case stdlib test suite.

Breaking

• Anything that invoked the deleted scraper scripts directly must switch to scripts/update.sh.

Install

Download professor-synapse.zip and upload it as a skill (or click "Copy to your skills"). The zip ships a clean memory seed — your store starts empty.

Updating from v1.1.0

Ask Professor Synapse to "check for updates," or run bash scripts/update.sh. It preserves your existing memory/ store byte-for-byte; only the code and protocols around it change.

Note: the .zip asset here is for humans — installed skills update via the codeload tarball of this tag, and version detection reads releases/latest.

v1.1.0 — Agent-tagged memory + ranked-fusion recall

Adds a persistent agent-tagged memory system, ranked-fusion recall, and a modernized versioned-update path. Backward-compatible with v1.0.0.

Highlights

• Persistent agent-tagged memory — shared store (scripts/memory.py): working memory (memory.json) + SQLite long-term store (longterm.db), every entry tagged with the agent that created it (recall broadly or filter by agent).
• Ranked-fusion recall — recall --query retrieves via FTS5 (stemming, prefix, column-weighted BM25) then re-ranks by fusing relevance + recency + record kind (Reciprocal Rank Fusion). New brief one-shot prefetch verb. No embeddings — fast and portable.
• 🧠 Memory Keeper agent + memory-protocol.md / memory-data-model.md.
• Agent summoning protocol extracted (summon-agent-protocol.md).
• Versioned codeload updates — the update protocol now pulls the canonical repo as a single codeload.github.com source tarball pinned to the release tag, preserving your custom agents, learned patterns, and memory/ store. Per-file HTML scraping demoted to a legacy fallback.
• 25-test suite (scripts/test_memory.py, stdlib unittest).

Install

Download professor-synapse.zip below → Claude → Settings → Capabilities → Skills → Add new skill.

Updating from v1.0.0

Ask Professor Synapse to "check for updates" — it compares its Version: to this tag and merges in via the codeload tarball, keeping your customizations and memory intact.

Full changelog: professor-synapse/references/changelog.md

v1.0.0 — Professor Synapse

Professor Synapse v1.0.0

First stable release of the Professor Synapse skill for Claude Desktop.

What's Included

• Expert agent orchestration — Summon domain specialists, create new agents on demand
• Domain Researcher — Built-in research agent for creating informed experts
• Convener Protocol — Multi-agent debate facilitation for complex decisions
• Two-tier learned patterns — Global patterns in SKILL.md, agent-specific patterns in each agent file
• Mandatory packaging workflow — Prominent enforcement so agents are never lost after creation
• Auto-healing index — rebuild-index.sh auto-appends Learned Patterns sections to agents missing them
• Changelog + Update Protocol — Targeted diffs when updating from upstream
• Self-updating skill — Learns and improves through pattern capture over time

Install

1. Download professor-synapse.zip from release assets
2. Claude → Settings → Capabilities → Skills → Add new skill
3. Upload the zip

GPT 2.0.0

Add donation link to README.md to support the project