Replace TinyAgents with TinyInference

.gitmodules

@@ -6,7 +6,6 @@
 	path = vendor/tinycortex
 	url = https://github.com/tinyhumansai/tinycortex.git
 	branch = main
-[submodule "vendor/tinyagents"]
-	path = vendor/tinyagents
-	url = https://github.com/tinyhumansai/tinyagents
-	branch = main
+[submodule "vendor/tinyinference"]
+	path = vendor/tinyinference
+	url = https://github.com/tinyhumansai/tinyinference.git

Cargo.toml

@@ -29,7 +29,8 @@ default-members = [
     "crates/tinymemory-sync",
     "crates/tinymemory-tinycortex",
 ]
-# `vendor/` holds engine submodules (tinycortex, tinybus, tinyagents), each of
+# `vendor/` holds pinned source submodules (tinycortex, tinybus,
+# tinyinference), each of
 # which is its own workspace with its own lockfile. Same exclusion
 # `vendor/tinycortex` uses for its own nested vendor directory.
 #
@@ -73,24 +74,19 @@ exclude = ["vendor", "crates/tinymemory-module", "worktrees"]
 [patch."https://github.com/tinyhumansai/tinymemory"]
 tinymemory-api = { path = "crates/tinymemory-api" }
 
+# TinyCortex pins TinyInference by git until it is published. Collapse that git
+# source onto the same checkout `tinymemory-core` resolves through crates.io so
+# both crates share one trait identity.
+[patch."https://github.com/tinyhumansai/tinyinference"]
+tinyinference = { path = "vendor/tinyinference/crates/tinyinference" }
+
 [patch.crates-io]
 tinycortex = { path = "vendor/tinycortex" }
 tinycortex-api = { path = "vendor/tinycortex/api" }
-# `tinymemory-core`'s summariser and embedder factory name tinyagents' chat and
-# embedding model traits. It is not published, so a standalone build of this
-# workspace has to be told where it is — hence its own submodule alongside
-# tinycortex and tinybus.
-#
-# Deliberately NOT `vendor/tinycortex/vendor/tinyagents`: the engine pins v2.1.0
-# there, which predates `RECOMMENDED_OLLAMA_CONTEXT_TOKENS` that the embedder
-# factory needs. Reaching through another submodule's pin also makes this
-# crate's dependency an accident of the engine's, which is not a relationship
-# worth having.
-#
-# A host that embeds this crate patches tinyagents itself — patch tables only
-# apply from the workspace root being built — so this entry affects standalone
-# builds and `cargo test` here, and nothing downstream.
-tinyagents = { path = "vendor/tinyagents" }
+# `tinymemory-core` names the provider-neutral chat and embedding contracts
+# from TinyInference. It is not published, so standalone builds resolve the
+# version requirement to the pinned checkout here.
+tinyinference = { path = "vendor/tinyinference/crates/tinyinference" }
 
 [profile.release]
 # Cross-crate optimization and smaller, faster binaries for release builds.

README.md

@@ -57,7 +57,7 @@ crates/
                         workspace on purpose — see the note in `Cargo.toml`.
 vendor/
 ├── tinycortex/         the engine, pinned as a submodule
-├── tinyagents/         pinned TinyAgents submodule
+├── tinyinference/      provider-neutral inference APIs, pinned as a submodule
 └── tinybus/            pinned TinyBus submodule
 ```
 
@@ -103,8 +103,8 @@ families answer — see the engine table under
 
 
 Run `git submodule update --init --recursive` after cloning. Nothing in the
-workspace builds without it — `tinymemory-core` names `tinyagents` and
-`tinycortex` by path through `vendor/`, so an uninitialized checkout fails at
+workspace builds without it — `tinymemory-core` names `tinyinference` and
+`tinycortex` by path through `vendor/`. An uninitialized checkout therefore fails at
 manifest resolution rather than at compile time, which reads as a confusing
 error.
 
@@ -134,7 +134,7 @@ without any `[patch]` entries.
 
 The remote recipe above works by git because the remote adapter reaches only
 published crates. The embedded engine does not: it pulls `tinycortex`,
-`tinycortex-api` and `tinyagents`, none of which are published, and
+`tinycortex-api` and `tinyinference`, none of which are published, and
 `tinycortex-api` takes `tinymemory-api` *by git*, which cargo will resolve as a
 second copy of a crate this workspace also provides by path. Patching that away
 needs the crates on disk, so the embedded path is a submodule dependency until
@@ -149,21 +149,25 @@ git -C vendor/tinymemory submodule update --init --recursive
 [dependencies]
 tinymemory = { path = "vendor/tinymemory", features = ["tinycortex"] }
 
-# All four are required. The first three are unpublished crates the engine
-# needs; the fourth collapses `tinycortex-api`'s git dependency on
-# `tinymemory-api` onto the copy in this tree — without it two distinct
+# All five entries are required. The three crates.io patches resolve unpublished
+# crates used by the memory layer and embedded engine. The TinyInference source
+# patch collapses TinyCortex's git dependency onto that same crate identity. The
+# fifth entry collapses `tinycortex-api`'s git dependency on `tinymemory-api`
+# onto the copy in this tree — without it two distinct
 # `tinymemory_api::MemoryEntry` types exist and the seam stops type-checking.
 [patch.crates-io]
 tinycortex = { path = "vendor/tinymemory/vendor/tinycortex" }
 tinycortex-api = { path = "vendor/tinymemory/vendor/tinycortex/api" }
-tinyagents = { path = "vendor/tinymemory/vendor/tinyagents" }
+tinyinference = { path = "vendor/tinymemory/vendor/tinyinference/crates/tinyinference" }
+[patch."https://github.com/tinyhumansai/tinyinference"]
+tinyinference = { path = "vendor/tinymemory/vendor/tinyinference/crates/tinyinference" }
 [patch."https://github.com/tinyhumansai/tinymemory"]
 tinymemory-api = { path = "vendor/tinymemory/api" }
 ```
 
 This exact patch set is what the reference consumer in
 `crates/tinymemory/examples/` and the repository's own root manifest use; a
-build missing any of the four fails at resolution, before compiling a line.
+build missing any of the five fails at resolution, before compiling a line.
 
 ```rust,ignore
 use std::sync::Arc;

crates/tinymemory-api/Cargo.toml

@@ -7,7 +7,7 @@ publish = false
 version = "0.1.1"
 edition = "2021"
 rust-version = "1.96"
-license = "MIT"
+license = "GPL-3.0-only"
 repository = "https://github.com/tinyhumansai/tinymemory"
 description = "Stable public contracts for the TinyMemory memory system"
 

crates/tinymemory-bus/Cargo.toml

@@ -6,7 +6,7 @@ publish = false
 version = "0.1.0"
 edition = "2021"
 rust-version = "1.96"
-license = "MIT"
+license = "GPL-3.0-only"
 repository = "https://github.com/tinyhumansai/tinymemory"
 description = "The TinyBus wire contract for the TinyMemory module: member names, payload types, and typed calls"
 

crates/tinymemory-conformance/Cargo.toml

@@ -4,7 +4,7 @@ publish = false
 version = "0.1.0"
 edition = "2021"
 rust-version = "1.96"
-license = "MIT"
+license = "GPL-3.0-only"
 description = "Behavioural conformance suite every MemoryProvider driver must pass"
 repository = "https://github.com/tinyhumansai/tinymemory"
 

crates/tinymemory-core/Cargo.toml

@@ -7,7 +7,7 @@ publish = false
 version = "0.1.0"
 edition = "2021"
 rust-version = "1.96"
-license = "MIT"
+license = "GPL-3.0-only"
 description = "The engine-neutral memory subsystem: store, summary tree, sync pipelines, ingestion and recall"
 repository = "https://github.com/tinyhumansai/tinymemory"
 readme = "../../README.md"
@@ -37,9 +37,10 @@ regex = "1.10"
 tinycortex = { version = "0.1", features = ["obsidian", "persona", "people", "sync"] }
 tinycortex-api = { version = "0.1" }
 
-# Chat-model and embedding primitives used by the tree summarizer and the
-# embedding factory.
-tinyagents = { version = "2.1", features = ["sqlite"] }
+# Provider-neutral chat-model and embedding primitives used by the tree
+# summarizer and embedding factory. Agent runtime and session behavior do not
+# belong in the memory layer.
+tinyinference = "0.2"
 anyhow = "1.0"
 async-trait = "0.1"
 # NO `git2` HERE, DELIBERATELY — do not add it back. `diff/` holds the ops and

crates/tinymemory-core/src/chat.rs

@@ -13,8 +13,8 @@ use async_trait::async_trait;
 
 use crate::chat_host::{create_chat_model_with_model_id, provider_for_role, UsageInfo};
 use crate::Config;
-use tinyagents::harness::message::Message;
-use tinyagents::harness::model::{ChatModel, ModelRequest};
+use tinyinference::message::Message;
+use tinyinference::model::{ChatModel, ModelRequest};
 
 /// One pair of prompt messages handed to the memory LLM backend.
 #[derive(Debug, Clone)]

crates/tinymemory-core/src/chat_host.rs

@@ -8,10 +8,10 @@
 //!
 //! # Why this trait is here and not in `tinymemory-api`
 //!
-//! It names `tinyagents::harness::model::ChatModel`, and the contract crate is
-//! deliberately dependency-light — it must not pull in tinyagents. This crate
-//! already depends on tinyagents, so it is the one place that can name both the
-//! model trait and the config seam. The host implements it here.
+//! It names [`tinyinference::model::ChatModel`], and the contract crate is
+//! deliberately dependency-light — it must not pull in an inference SDK. This
+//! crate already depends on TinyInference, so it is the one place that can name
+//! both the model trait and the config seam. The host implements it here.
 //!
 //! Reached through a process-global for the same reason as
 //! [`crate::embedding_host`]; see that module for the rationale, and for why an
@@ -20,7 +20,7 @@
 use std::sync::Arc;
 
 use parking_lot::RwLock;
-use tinyagents::harness::model::{ChatModel, ModelResponse};
+use tinyinference::model::{ChatModel, ModelResponse};
 
 use crate::Config;
 

crates/tinymemory-core/src/embedding_adapter.rs

@@ -1,24 +1,24 @@
-//! [`TinyAgentsEmbeddingProvider`] — the adapter from tinyagents' own embedding
+//! [`TinyInferenceEmbeddingProvider`] — the adapter from TinyInference's embedding
 //! model trait onto the seam's [`EmbeddingProvider`].
 //!
 //! It lives in this crate rather than in `tinymemory-api` because the contract
-//! crate must stay dependency-light and cannot name `tinyagents`; and rather
+//! crate must stay dependency-light and cannot name `tinyinference`; and rather
 //! than in the host because the tree's embedder factory — which is core code —
 //! builds Ollama models directly and needs to wrap them. The host re-exports it
 //! from `inference::embeddings`, so every existing path there keeps resolving
 //! and keeps naming this one type.
 
 use async_trait::async_trait;
-use tinyagents::harness::embeddings::EmbeddingModel;
+use tinyinference::embeddings::EmbeddingModel;
 
 pub use tinymemory_api::host::{format_embedding_signature, EmbeddingProvider};
 
-/// Compatibility adapter from the canonical tinyagents embedding model.
-pub struct TinyAgentsEmbeddingProvider {
+/// Compatibility adapter from the canonical TinyInference embedding model.
+pub struct TinyInferenceEmbeddingProvider {
     model: Box<dyn EmbeddingModel>,
 }
 
-impl TinyAgentsEmbeddingProvider {
+impl TinyInferenceEmbeddingProvider {
     pub fn new(model: impl EmbeddingModel + 'static) -> Self {
         Self {
             model: Box::new(model),
@@ -31,7 +31,7 @@ impl TinyAgentsEmbeddingProvider {
 }
 
 #[async_trait]
-impl EmbeddingProvider for TinyAgentsEmbeddingProvider {
+impl EmbeddingProvider for TinyInferenceEmbeddingProvider {
     fn name(&self) -> &str {
         self.model.name()
     }

crates/tinymemory-core/src/engine/mod.rs

@@ -3,7 +3,7 @@
 //!
 //! OpenHuman's memory subsystem migrates onto the `tinycortex` crate (store /
 //! chunks / tree / retrieval / queue / ingest / score + the long tail). This
-//! module is the **adapter seam**, mirroring `src/openhuman/agent/tinyagents/`: it
+//! module is the **adapter seam**, mirroring the host's inference adapters: it
 //! implements the crate's engine traits over OpenHuman services and derives the
 //! engine's [`tinycortex::memory::MemoryConfig`] from the host `Config`. Nothing here contains
 //! engine logic — that lives in the crate.

crates/tinymemory-core/src/sources/reconcile.rs

@@ -119,19 +119,6 @@ pub async fn apply_composio_source_caps_migration() -> Result<(), String> {
     Ok(())
 }
 
-#[cfg(test)]
-fn short_id(id: &str) -> &str {
-    // Show only the last 8 Unicode scalar values to keep labels compact.
-    // Byte-slicing would panic if the cut point isn't a UTF-8 boundary.
-    let n = id.chars().count();
-    if n <= 8 {
-        return id;
-    }
-    let skip = n - 8;
-    let start = id.char_indices().nth(skip).map(|(idx, _)| idx).unwrap_or(0);
-    &id[start..]
-}
-
 #[cfg(test)]
 #[path = "reconcile_tests.rs"]
 mod tests;

crates/tinymemory-core/src/sources/reconcile_tests.rs

@@ -3,6 +3,18 @@
 use super::*;
 use crate::sources::types::{MemorySourceEntry, SourceKind};
 
+fn short_id(id: &str) -> &str {
+    // Show only the last 8 Unicode scalar values to keep labels compact.
+    // Byte-slicing would panic if the cut point isn't a UTF-8 boundary.
+    let n = id.chars().count();
+    if n <= 8 {
+        return id;
+    }
+    let skip = n - 8;
+    let start = id.char_indices().nth(skip).map(|(idx, _)| idx).unwrap_or(0);
+    &id[start..]
+}
+
 fn make_composio_entry(
     id: &str,
     toolkit: &str,

crates/tinymemory-core/src/store/factories.rs

@@ -18,7 +18,7 @@ use rusqlite::Connection;
 use crate::embedding_host::require_embedding_host;
 use crate::store::namespace_store::UnifiedMemory;
 use crate::traits::Memory;
-use tinyagents::harness::embeddings::{DEFAULT_OLLAMA_DIMENSIONS, DEFAULT_OLLAMA_MODEL};
+use tinyinference::embeddings::{DEFAULT_OLLAMA_DIMENSIONS, DEFAULT_OLLAMA_MODEL};
 use tinymemory_api::host::MemoryConfig;
 use tinymemory_api::host::{format_embedding_signature, EmbeddingProvider};
 use tinymemory_api::host::{EmbeddingRouteConfig, StorageProviderConfig};