Mnemonic Proxy brings archive recall to Hermes and patches a false-finish trap
The local-model proxy's new guide connects Hermes without a plugin; its latest fix catches recall markup that could otherwise pass as a successful final answer.
Hermes users running local models have a new documented way to keep old tool results retrievable after they leave the prompt. Mnemonic Proxy’s v0.3.18 release adds a Hermes Agent setup guide and fixes a failure mode in which a model’s attempted recall could masquerade as a finished answer.
Mnemonic is a standalone HTTP proxy, not a Hermes plugin or a new model. It sits between the harness and an OpenAI-compatible backend, archives messages and tool results verbatim, and replaces older prompt material with receipts. Its own recall tool retrieves archived material internally.
Point Hermes at the proxy
The version-pinned guide uses Hermes’ custom Chat Completions provider, a base URL such as http://127.0.0.1:8096/v1, and compression.enabled=false. Disabling Hermes’ separate compression lets the proxy receive the original history and manage masking and recall itself.
The README documents installation with pip install mnemonic-proxy on Python 3.10 or newer, followed by its demo and engine checks. Model serving remains a separate requirement; the proxy does not supply a hosted model or inference credits. Linux is the reference platform. Native Windows support has CI testing but no reported desktop validation, and macOS remains untested.
The guide recommends an isolated Hermes profile. A dummy key is appropriate only if the local endpoint requires no authentication; it explicitly warns against reusing a cloud-provider key. Hermes’ normal tool approvals still apply, and the proxy’s disk archive contains the original conversation.
What the release actually establishes
The author reports an October 28 integration run with Hermes 0.22.2 and Mnemonic Proxy 0.3.17, using Qwen3.8 Flash Next through Strata. That run handled 18 user turns, 24 client tool calls and one archive recall through a deliberately reduced 32K proxy window. These are the project’s results, not an AITrending reproduction or a guarantee for other models.
The new 0.3.18 fix concerns a different failure: after the proxy removed its recall tool declaration, Strata could return a residual recall call as plain text. The proxy previously accepted that text as a successful final answer. The release routes it through retry and explicit failure-notice handling for streaming and non-streaming responses.
The maintainer says the fix was checked on CPU using a real response as a fixture, with a real-model rerun still pending. The guide also records a failed oversized-tool-output probe. Archive masking does not enlarge the underlying engine’s actual context capacity or make an oversized current result fit automatically.