Synapse Layer Skill for Hermes Agent

[rafacpti23]

This skill integrates Synapse Layer — a zero-knowledge, encrypted, cross-session memory system — into Hermes Agent. All memories are encrypted with AES-256-GCM, pass through a 4-layer Cognitive Security Pipeline (PII redaction, differential privacy, intent validation, neural handover), and are retrievable from any channel (Telegram, WhatsApp, CLI, Discord).

Credits

Rafa Martins — synapselayer.org

What It Does

• Persistent Memory: Context survives between sessions. You don't repeat yourself.
• Cross-Channel: Telegram, WhatsApp, CLI, and Discord all share the same memory pool.
• Trust Quotient (TQ): Every memory gets a confidence score (0.0–1.0) so the agent knows what to trust.
• Zero-Knowledge Encryption: Memories are encrypted before leaving the agent. The server never sees plaintext.
• PII Auto-Redaction: Emails, phone numbers, API keys, IPs are automatically stripped before storage.
• Differential Privacy: Noise is injected into embeddings to prevent inference attacks

[teknium1]

Thanks for documenting the Synapse Layer integration. The current patch needs substantive rework before it can be safely used.

Problems

• skills/synapse-layer/SKILL.md:160-165 tells the agent to save API credentials and tokens to memory. Hermes explicitly prohibits preserving API keys, tokens, passwords, and credentials (agent/context_compressor.py:1804-1806).
• skills/synapse-layer/SKILL.md:5 has a 203-character description; AGENTS.md:888-900 requires a <=60-character sentence ending in a period. The standard frontmatter fields are also absent.
• skills/synapse-layer/SKILL.md:64 documents hermes restart; current main exposes hermes gateway restart (hermes_cli/subcommands/gateway.py:120-123) and documents /reload-mcp for MCP reloads (website/docs/reference/faq.md:568-576).
• skills/synapse-layer/SKILL.md:131-187 contains personal infrastructure and hardcoded /root/.hermes paths, so it is not reusable or profile-safe.

Suggested changes

• Remove all credential-persistence instructions.
• Rework the skill to current frontmatter, section, category, and test requirements.
• Use verified MCP setup/reload instructions and generic examples. If automatic memory lifecycle is intended, implement it as a standalone MemoryProvider plugin rather than claiming it from Markdown alone.

Automated hermes-sweeper review.

[teknium1]

Current skill standards require a one-sentence description of at most 60 characters ending in a period (AGENTS.md:888-900); this description is 203 characters. Please shorten it and add the standard author, license, platforms, and metadata.hermes frontmatter.

[teknium1]

Do not instruct the agent to store credentials or tokens in external memory. Remove this category and every related save instruction; Hermes requires API keys, tokens, passwords, and credentials to be redacted rather than preserved (agent/context_compressor.py:1804-1806).

[GottZ]

This was generated by AI during triage.

Summary

Nine PRs address or reference persistent-memory integrations, but their diffs implement materially different backends or integration shapes: provider plugins, a core SQLite replacement, an external toolset, and—only in #11678—a Synapse Layer skill. The target diff documents Synapse setup rather than changing the memory runtime, and its current documentation contains security, schema, command, and portability defects identified by contributor review.

Related pull requests

• #1811 [closed] related — (+1366/-9) — reference only: Adds a full Hindsight cloud/local integration with CLI setup, runtime hooks, and retain/recall/reflect tools, rather than addressing the Synapse skill documentation. Although closed, it remains relevant as an approved reference implementation for a substantially deeper memory integration.
• #4480 [closed] related — (+6073/-163) — do not consolidate: Replaces flat-file memory with a large SQLite/FTS5 knowledge engine, extraction, graph, lifecycle, and consolidation subsystems; it does not repair or supersede the Synapse skill. Although closed, it remains relevant as an architectural reference, but the author retracted it after identifying dead code and unexercised subsystems.
• #5671 [closed] related — (+1487/-0) — superseded duplicate outside this set: Adds an in-tree MemPalace provider with semantic recall, lifecycle hooks, and tools, not Synapse documentation. Although closed, it remains relevant because contributor discussion identified it as a duplicate of #12203, and the author later closed it in favor of the canonical rebased successor #21017; the reported embedding-dimension mismatch also blocked automatic turn persistence in this version.
• #5855 [closed] related — (+940/-0) — separate implementation: Adds an in-tree ALIVE file-based context provider with walnut discovery, smart prefetch, persistence hooks, and three tools; it neither duplicates nor fixes #11678. Although closed, it remains relevant as a contrasting local structured-context design, not as a merge candidate for this skill.
• #6114 [closed] related — (+126/-0) — separate implementation: Integrates the external deep-memory package through optional dependencies, tool discovery, prompt injection, and post-session reasoning; it does not address Synapse skill correctness. Although closed, it remains relevant as a lightweight external-package integration reference.
• #11678 related — (+437/-0) — keep open, do not merge yet: Adds only Synapse Layer README/SKILL documentation and MCP setup, but the diff instructs agents to persist credentials, uses noncompliant frontmatter, documents obsolete hermes restart, and embeds personal /root/.hermes infrastructure. This follows the visible contributor keep_open review on #11678: salvage requires removing credential-persistence instructions, adopting current skill frontmatter, using hermes gateway restart or /reload-mcp as appropriate, and replacing personal/profile-unsafe content with reusable profile-aware guidance.
• #20005 [closed] related — (+672/-0) — standalone-plugin path: Adds an in-tree Muninn semantic MemoryProvider with HTTP routing, peer activation, prefetch, and persistence, not a Synapse skill. Although closed, it remains relevant because contributor policy explicitly rejects new in-tree memory providers and directs this implementation to a standalone plugin plus an optional community-docs PR.
• #33379 [closed] related — (+2777/-0) — duplicate of #33403: Adds the first in-tree enhanced-memory implementation with two-tier SQLite/FTS5 storage, condensation, embeddings, and semantic search; it does not address #11678. Although closed, it remains relevant as the earlier version of the same implementation later expanded in #33403 and moved toward a standalone repository under current policy.
• #33403 [closed] related — (+7319/-0) — superseding duplicate, standalone only: Expands #33379's same enhanced-memory implementation with tests and documentation, while remaining unrelated to the Synapse skill defects. Although closed, it remains relevant as the fuller version; the author closed it under the standalone-memory-provider policy and moved the work to hermes-enhanced-memory.

Duplicates

#33379 and #33403 are substantially the same enhanced-memory implementation, with #33403 expanding #33379. Separately, contributor discussion identifies #5671 as a duplicate of out-of-set #12203, and it was ultimately closed in favor of out-of-set canonical successor #21017.

Suggested consolidation

Do not merge #11678 in its current form; retain it as the sole in-scope Synapse candidate, consistent with the contributor keep_open review, and require the diff itself to remove all credential-persistence guidance, use valid current skill frontmatter, replace the obsolete restart command, and eliminate personal or hardcoded profile-unsafe content before re-review. Do not reopen or merge the closed provider PRs into this consolidation: #33379 can remain closed as a duplicate of #33403, #33403 and #20005 belong on the standalone-plugin path, #5671 is superseded by #12203/#21017, and #1811, #4480, #5855, and #6114 are distinct reference implementations rather than duplicates of #11678.

Cross-PR triage: Reviewed 9 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 162 kB of PR diffs, 27 kB of issue/PR text, 7 kB of discussion (19 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[rafacpti23]

ok, avaliation ok